Article

Ensemble lightning prediction models for the province of Alberta, Canada

Authors:
  • Environment and Climate Change Canada, Edmonton, Canada
To read the full-text of this research, you can request a copy directly from the authors.

Abstract

Lightning is a major cause of wildland fires in Canada. During an average year in the province of Alberta, 330000 cloud-to-ground lightning strikes occur. These strikes are responsible for igniting 45% of reported wildfires (∼450 fires) and 71% of area burned (∼105000ha). Lightning-caused wildland fires in remote areas have large suppression costs and a greater chance of escaping initial attack when compared with human-caused fires, which are often located close to infrastructure and suppression resources. In this study, geographic and temporal covariates were paired with meteorological reanalysis and radiosonde observations to generate a series of 6-h and 24-h lightning prediction models valid from April to October. These models, based on cloud-to-ground lightning from the Canadian Lightning Detection Network, were developed and validated for the province of Alberta, Canada. The ensemble forecasts produced from these models were most accurate in the Rocky Mountain and Foothills Natural Regions, achieving hits rates of 85%. The Showalter index, latitude, elevation, longitude, Julian day and convective available potential energy were found to be highly important predictors. Random forest classification is introduced as a viable modelling method to generate lightning forecasts.

No full-text available

Request Full-text Paper PDF

To read the full-text of this research,
you can request a copy directly from the authors.

... al. (Mostajabi et al., 2019) used XGBoost to perform 30 minutes ahead forecasting of lightning occurrence based on a set of single-site observations of meteorological parameters. (Blouin et al., 2016) used a tree-based classification algorithm to predict 6-hours and 24hours CG lightning. A 1-hour nowcasting model is proposed in (Mecikalski et al., 2015), showing that lightning forecast are made 45 minutes before rainfall occurs. ...
... The classification has been performed using RF. The choice is due to the demonstrated potential and benefits in the prediction ability of the RF technique for nowcasting problems strictly correlated with the lightning phenomenon, such as the prediction of small-scale storm initiation, diagnosis of turbulence, mesoscale convective system initiation and lightning activity, respectively shown in (Breiman, 2001) (Williams, 2014) (Blouin et al., 2016) (Ahijevych et al., 2016). RF is an ensemble method based on Classification and Regression Decision Trees, originally introduced in (Breiman, 2001). ...
... Results reached confirm the yet high POD and low FAR obtained in (Blouin et al., 2016), in which a lower spatial resolution was used, and in (Mostajabi et al., 2019), in which the best results were obtained with an XGBoost algorithm to perform 30 minutes ahead forecasting of lightning occurrence based on a set of single-site observations of meteorological parameters. Differently, results in (Zhou et al., 2020), obtained using data from geostationary meteorological satellites as input to create a DL algorithm, reach higher performances with respect to the presented model when dealing with hours characterized by high intense lightning activity. ...
Preprint
Full-text available
This paper discusses the use of Random Forest (RF), a popular Machine Learning (ML) algorithm, to perform spatially explicit nowcasting of cloud-to-ground lightning occurrence. An application to the Italian territory and the surrounding seas is presented. Specifically, 1-hour ahead lightning occurrences over the months of August, September and October from 2017 to 2019 have been modelled using a dataset including geo-environmental features. Results obtained with three different spatial resolutions have been compared, for nowcasting both positive and negative strokes. The features’ importance resulting from the best RF models showed how datadriven models are able to identify the relationships between meteorological variables, in agreement with previous physically based knowledge of the phenomenon. The encouraging results obtained in terms of forecasting accuracy support the idea to use ML-based algorithms in early warning procedures for disaster risk management.
... Mostajabi et al. (2019) used XGBoost to perform 30 min ahead forecasting of lightning occurrence based on a set of single-site observations of meteorological parameters. (Blouin et al. 2016) used a tree-based classification algorithm to predict 6-h and 24-h CG lightning. A 1-h nowcasting model is proposed in (Mecikalski et al. 2015), showing that lightning forecast are made 45 min before rainfall occurs. ...
... The classification has been performed using RF. The choice is due to the demonstrated potential and benefits in the prediction ability of the RF technique for nowcasting problems strictly correlated with the lightning phenomenon, such as the prediction of small-scale storm initiation, diagnosis of turbulence, mesoscale convective system initiation and lightning activity, respectively shown in (Breiman 2001;Williams 2014;Blouin et al. 2016;Ahijevych et al. 2016). RF is an ensemble method based on Classification and Regression Decision Trees, originally introduced in (Breiman 2001). ...
... Distinction in the results obtained for positive and negative strokes may be also attributed to the differences in the creation mechanisms (Nag and Rakov 2012) and (Cooray and Arevalo 2017). Results reached confirm the yet high POD and low FAR obtained in (Blouin et al. 2016), in which a lower spatial resolution was used, and in (Mostajabi et al. 2019), in which the best results were obtained with an XGBoost algorithm to perform 30 min ahead forecasting of lightning occurrence based on a set of single-site observations of Yellow bars refer to the model created re-gridding all data of the DF over the HW resolution. Green bars refer to the model created regridding all data of the DF over the HT resolution Fig. 3 Features' importance of the resulting best RF model obtained for data gridded with HW resolution, for positive (left) and negative (right) strokes. ...
Article
Full-text available
The relation between the increase in the frequency and the effects of extreme events with climate change has been widely demonstrated and the related consequences are a global concern. In this framework, the strong correlation between significant lightning occurrence and intense precipitation events has been also documented. Consequently, the possibility of having a short-term forecasting tool of the lightning activity may help in identifying and monitoring the evolution of severe weather events on very short time ranges. The present paper proposes an application of Random Forest (RF), a popular Machine Learning (ML) algorithm, to perform a nowcasting of Cloud-to-Ground (CG) lightning occurrence over the Italian territory and the surrounding seas during the months of August, September, and October from 2017 to 2019. Results obtained with three different spatial resolutions have been compared, suggesting that, to enhance the skills of the model in identifying the presence or absence of strokes, all the data selected as input should be commonly gridded on the finest available spatial resolution. Moreover, the features’ importance analysis performed confirms that meteorological features describing the state of the atmosphere, especially at higher altitudes, have a stronger impact on the final result than topology data, such as Latitude or Digital Elevation Model (DEM).
... Previously, Stocks et al. (2002) examined national fire records of large fires in Canada and revealed a gradient from predominantly human-ignited fires in the more populous south of the country to predominantly lightning-ignited fires in the more remote northern regions. Around 70% of large fires and 85% of large fire BA were associated with lightning during 1959-1997 (Blouin et al., 2016;Stocks et al., 2002). Peterson et al. (2010) found that annual fire counts correlated strongly with lighting strike frequency and that ∼40% of satellite-detected hotspots in Interior Alaska were likely caused by lightning strikes during 2000-2006, while Kasischke et al. (2010) showed that lightning-ignited fire BA has increased since the 1940s in contrast to a decrease in the case of human-ignited fires. ...
... and satellite detections (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015). Numerous studies have also highlighted that climatic and topographic conditions are favorable for dry lightning occurrence in the North American boreal region (Blouin et al., 2016;Dissing & Verbyla, 2003;Peterson et al., 2010). The distribution of fires across populated and unpopulated regions of Russia infers that forest fires in Siberia are predominantly caused by lightning (Kharuk et al., 2011;Mollicone et al., 2006;Ponomarev et al., 2016), although the explicit separation of human and lightning -ignited fires is pending in the region. ...
... Regional assessments have similarly shown that slope and aspect can influence the occurrence and spread of fires in some regions with complex terrain, either because the topographical variation itself affects the dynamics of energy transfer or because topographical variation is a proxy for microclimatic conditions or ignition opportunities (Blouin et al., 2016;Cavard et al., 2015;Dillon et al., 2011;Fang et al., 2015;Iniguez et al., 2008;Oliveira et al., 2014;P. J. Clarke et al., 2014). ...
Article
Full-text available
Recent wildfire outbreaks around the world have prompted concern that climate change is increasing fire incidence, threatening human livelihood and biodiversity, and perpetuating climate change. Here, we review current understanding of the impacts of climate change on fire weather (weather conditions conducive to the ignition and spread of wildfires) and the consequences for regional fire activity as mediated by a range of other bioclimatic factors (including vegetation biogeography, productivity and lightning) and human factors (including ignition, suppression, and land use). Through supplemental analyses, we present a stocktake of regional trends in fire weather and burned area (BA) during recent decades, and we examine how fire activity relates to its bioclimatic and human drivers. Fire weather controls the annual timing of fires in most world regions and also drives inter‐annual variability in BA in the Mediterranean, the Pacific US and high latitude forests. Increases in the frequency and extremity of fire weather have been globally pervasive due to climate change during 1979–2019, meaning that landscapes are primed to burn more frequently. Correspondingly, increases in BA of ∼50% or higher have been seen in some extratropical forest ecoregions including in the Pacific US and high‐latitude forests during 2001–2019, though interannual variability remains large in these regions. Nonetheless, other bioclimatic and human factors can override the relationship between BA and fire weather. For example, BA in savannahs relates more strongly to patterns of fuel production or to the fragmentation of naturally fire‐prone landscapes by agriculture. Similarly, BA trends in tropical forests relate more strongly to deforestation rates and forest degradation than to changing fire weather. Overall, BA has reduced by 27% globally in the past two decades, due in large part to a decline in BA in African savannahs. According to climate models, the prevalence and extremity of fire weather has already emerged beyond its pre‐industrial variability in the Mediterranean due to climate change, and emergence will become increasingly widespread at additional levels of warming. Moreover, several of the major wildfires experienced in recent years, including the Australian bushfires of 2019/2020, have occurred amidst fire weather conditions that were considerably more likely due to climate change. Current fire models incompletely reproduce the observed spatial patterns of BA based on their existing representations of the relationships between fire and its bioclimatic and human controls, and historical trends in BA also vary considerably across models. Advances in the observation of fire and understanding of its controlling factors are supporting the addition or optimization of a range of processes in models. Overall, climate change is exerting a pervasive upwards pressure on fire globally by increasing the frequency and intensity of fire weather, and this upwards pressure will escalate with each increment of global warming. Improvements to fire models and a better understanding of the interactions between climate, climate extremes, humans and fire are required to predict future fire activity and to mitigate against its consequences.
... Lifting conditions are usually provided by meso-and smallscale systems; moisture and instability conditions are related to the development and evolution of synoptic-scale systems. Fortunately, the machine learning model can well describe the nonlinear relationship between the occurrence of deep convection and the convective parameters (Wang et al., 2014;Collins and Tissot, 2015;Blouin et al., 2016;Ukkonen et al., 2017;Ukkonen and Mäkelä, 2019), and is conducive to reproduce the geographical distribution and diurnal variation of thunderstorm frequency. Notably, the same model is not suitable for different regions due to the influence of geographical location and atmospheric circulation (Gasc on et al., 2015). ...
... Compared with multiple statistical regression (Dai et al., 2009), decision trees (Burrows et al., 2005) and neural networks (Manzato, 2007), RF can model complex interactions, deal with imbalanced data and rank the importance of variables. RF has been widely used in research on weather forecasting (Ahijevych et al., 2016;Blouin et al., 2016;Czernecki et al., 2019). In consideration of the situation discussed above, nonlinear RF combined with unbalanced data down-sampling and optimal feature selection strategies were used to create ML models for thunderstorm simulation. ...
Article
Full-text available
Retrieving thunderstorm activity through specific thermodynamic and kinematic parameters is paramount for predicting deep convective weather and investigating long‐term climatology of storm. However, the reliability of the relationship between parameters and convective events is restricted by the modelling methods and sampling of thunderstorm activity. There is no objective definition of a thunderstorm, so the clustering method is applied to the cloud‐to‐ground (CG) lightning stroke data in Central China to identify lightning clusters. These clusters are then gridded and associated with environmental variables derived from ERA5 reanalysis. Finally, machine learning (ML) technologies are applied to model the occurrence of thunderstorms. In addition, ERA5 is also evaluated. The parameters related to moisture and lapse rate calculated by ERA5 are close to sounding measurements, such as Td850, PW, LR700_400 and KI, whose correlation coefficients exceed 0.90. ERA5 has a good estimation of some parameters that are susceptible to the influence of the boundary layer. Compared with the lightning strike‐based scheme, our scheme obtains the best performance index values. An agreement between observations and predictions based on lightning clusters is also evident from the diurnal cycle of thunderstorm probabilities. Although thunderstorm activity on complex terrain is underestimated, the created ML model can explain 61.4% of the variance in the observed frequency. The results of significance test reveal that there are statistically significant differences between the soundings corresponding to some isolated CG strikes and the thunderstorm class, but the distribution is the same as that of the non‐thunderstorm class. Solar radiation, topographic features and lake distribution play a major role in promoting the occurrence and development of thunderstorms. This article is protected by copyright. All rights reserved.
... The remaining six years were used to validate the model. A detailed look at random forest and the methodology used to generate the models is available in Blouin (2014) and Blouin et al. (2016). ...
... The study was published in the International Journal of Wildland Fire (Blouin et al., 2016) and was funded by the Western Partnership for Wildland Fire Science. ...
... For instance, lightning is predicted to increase by approximately 50% over the 21st century in the contiguous USA (Romps et al. 2014). In Canada, lightning is already a major cause of wildfire in many areas (Blouin et al. 2016;Stocks et al. 2002;Wierzchowski et al. 2002), responsible for about 50% of large fires and 90% of area burned (Hanes et al. 2019). A study of the northern boreal forest found that the large amount of area burned in 2014 and 2015 coincided with a record amount of lightning strikes (Veraverbeke et al. 2017). ...
... For example, there has been an increased use of machine learning and artificial intelligence (ML-AI) techniques in wildfire science covering a range of topics. Examples include extreme fire weather prediction (Lagerquist et al. 2017), ensemble lightning prediction modelling (Blouin et al. 2016), and automated burn scar mapping (Cao et al. 2009). Other advancements in fire science worth mentioning include improved numerical weather and fire growth models. ...
Article
Full-text available
Recently, the World Scientists’ Warning to Humanity: a Second Notice was issued in response to ongoing and largely unabated environmental degradation due to anthropogenic activities. In the warning, humanity is urged to practice more environmentally sustainable alternatives to business as usual to avoid potentially catastrophic outcomes. Following the success of their warning, the Alliance of World Scientists called for discipline-specific follow-up papers. This paper is an answer to that call for the topic of wildland fire. Across much of Canada and the world, wildfires are anticipated to increase in severity and frequency in response to anthropogenic activities. The world scientists’ second warning provides the opportunity for wildland fire researchers to raise the profile of the potential impacts that anthropogenic activities are likely to have on future fire regimes and, in return, what impacts future fire regimes may have on humanity. We discuss how wildfire is related to several issues of concern raised in the world scientists’ second warning, including climate change, human population growth, biodiversity and forests, and freshwater availability. Furthermore, we touch on the potential future health impacts and challenges to wildfire suppression and management in Canada. In essence, our wildfire scientists’ warning to humanity is that we, as a society, will have to learn to live with more fire on the landscape. We provide some recommendations on how we might move forward to prepare for and adapt to future wildfire regimes in Canada. Although this paper is primarily Canadian in focus, the concepts and information herein also draw from international examples and are of relevance globally.
... Lately, the ensemble methods have gained significant research attention as they are known to produce better performance than the traditional classifiers on a number of problems [3,8,20,32,35,52,[56][57][58]. Likewise, in Customer Churn Prediction domain, various ensembles have been employed, giving better results than individual classifiers. ...
Article
Full-text available
Combining multiple classifiers to create hybrid learners (ensembles) has gained popularity in recent years. Ensembles are gaining more interest in the field of data mining as they have reportedly performed best predictions as compared to individual classifiers. This has resulted in experimentation with new ways of ensemble creation. This paper presents a study on creation of novel hybrid ways of combining multiple ensemble models using ‘over production and choose approach.’ In contrast to the original concept of ensembles that combine various learners, the proposed ensemble models comprise of combinations of other ensembles. In particular, we have combined learners as in composition of other learners, thus producing nested learners. Two such models named as Boosted-Stacked learners and Bagged-Stacked learners are proposed and are shown to outperform the traditional ensembles. Experiments are performed in churn prediction domain where a benchmark customer churn dataset (available on UCI repository) and a newly created dataset from a South Asian wireless telecom operator (named as SATO) are used. SATO dataset is created as balanced dataset (having equal number of churners and non-churners). The novel Boosted-Stacked learner and Bagged-Stacked learner achieved accuracies of 98.4% and 97.2%, respectively, on the UCI Churn dataset outperforming the existing state-of-the-art techniques. Furthermore, a high accuracy on the SATO dataset validates the effectiveness of the proposed models on balanced as well as imbalanced datasets.
... Appreciable attention has been given to the problems of thunderstorm classification and prediction in recent years. Many techniques have been used, including empirical orthogonal function and canonical correlation analyses (e.g., Muñoz et al. 2016), classification and regression trees (e.g., Burrows et al. 2005), random-forest classification (e.g., Blouin et al. 2016), quadratic discriminant analysis (e.g., Sánchez et al. 1998), logistic regression (e.g., Mazany et al. 2002;Sousa et al. 2013;Romps et al. 2014), and dynamical modeling (e.g., Yair et al. 2010;Lynn et al. 2012;Zepka et al. 2014). Subject areas have included very-short-range forecasting, seasonal prediction, and climatological studies. ...
Article
Lightning is a natural hazard that can lead to the ignition of wildfires, disruption and damage to power and telecommunication infrastructures, human and livestock injuries and fatalities, and disruption to airport activities. This paper examines the ability of six statistical and machine-learning classification techniques to distinguish between nonlightning and lightning days at the coarse spatial and temporal scales of current general circulation models and reanalyses. The classification techniques considered were 1) a combination of principal component analysis and logistic regression, 2) classification and regression trees, 3) random forests, 4) linear discriminant analysis, 5) quadratic discriminant analysis, and 6) logistic regression. Lightning-flash counts at six locations across Australia for 2004-13 were used, together with atmospheric variables from the ERA-Interim dataset. Tenfold cross validation was used to evaluate classification performance. It was found that logistic regression was superior to the other classifiers considered and that its prediction skill is much better than using climatological values. The sets of atmospheric variables included in the final logistic-regression models were primarily composed of spatial mean measures of instability and lifting potential, along with atmospheric water content. The memberships of these sets varied among climatic zones.
... Modelling human-caused wildfire ignitions. Recent studies have used lightning ignition data to estimate wildfire risk during the wildfire season 27,28 . However, wildfire risk is known to vary according to ignition source and season 29 . ...
Article
Full-text available
Spring represents the peak of human-caused wildfire events in populated boreal forests, resulting in catastrophic loss of property and human life. Human-caused wildfire risk is anticipated to increase in northern forests as fuels become drier, on average, under warming climate scenarios and as population density increases within formerly remote regions. We investigated springtime human-caused wildfire risk derived from satellite-observed vegetation greenness in the early part of the growing season, a period of increased ignition and wildfire spread potential from snow melt to vegetation green-up with the aim of developing an early warning wildfire risk system. The initial system was developed for 392,856 km² of forested lands with satellite observations available prior to the start of the official wildfire season and predicted peak human-caused wildfire activity with 10-day accuracy for 76% of wildfire-protected lands by March 22. The early warning system could have significant utility as a cost-effective solution for wildfire managers to prioritize the deployment of wildfire protection resources in wildfire-prone landscapes across boreal-dominated ecosystems of North America, Europe, and Russia using open access Earth observations.
... We will mention a few details about the random forest procedure here, but space does not allow us to provide more due to the algorithm's complexity. We recommend the interested reader consult the literature for details, for example Blouin et al. (2016), Rodriguez-Galiano et al. (2012), Cutler et al. (2007). Whereas CART builds a single decision tree from learning data, random forest constructs a group of uncorrelated decision trees by bootstrapping the learning data and constructing a decision tree with each bootstrapped data sample. ...
Article
Blizzard conditions occur regularly in the Canadian Arctic, with high impact on travel and life there. These extreme conditions are challenging to forecast for this vast domain because the observation network is sparse and remote sensing coverage is limited. To establish occurrence statistics we analyzed METeorological Aerodrome Reports (METARs) from Canadian Arctic stations between October and May 2014-2018. Blizzard conditions occur most frequently in open tundra east and north of the boreal forest boundary, with highest frequency found on the northwest side of Hudson Bay and over flat terrain in central Baffin Island. Except in sheltered locations, the reported cause of reduced visibility is blowing snow without precipitating snow in about one-half to two-thirds of METARs made by a human observer, even higher at some stations. We produce three products that forecast blizzard conditions from post-processed NWP model output. The blizzard potential (BP), generated from expert’s rules, is intended for warning well in advance of areas where blizzard conditions may develop. A second product (BH) stems from regression equations for the probability of visibility ≤ 1 km in blowing snow and/or concurrent snow derived by Baggaley and Hanesiak (2005). A third product (RF), generated with the Random Forest ensemble classification algorithm, makes a consensus YES/NO forecast for blizzard conditions. We describe the products, provide verification, and show forecasts for a significant blizzard event. Receiver Operator Characteristic curves and critical success index scores show RF forecasts have greater accuracy than BP and BH forecasts at all lead times.
... Comparisons between statistical and algorithmic procedures have revealed that both techniques show acceptable levels of predictive ability of fire ignitions, although algorithmic models present a better accuracy and robustness (Bar-Massada et al. 2013). Overall, human-and lightning-caused ignitions are usually considered separately (as seen in Wang and Anderson 2011), given that the underlying processes generating the ignitions are distinct and lead to unique spatial and temporal patterns (Flannigan and Wotton 1991;Krawchuk et al. 2006;Morissette and Gauthier 2008;Blouin et al. 2016;Coughlan et al. 2018). ...
Article
Full-text available
Despite increasing concern about wildland fire risk in Canada, there is little synthesis of knowledge that could contribute to the development of a comprehensive risk framework for a wide range of values, which is an essential need for the country. With dramatic variability in costs and losses from this natural hazard, there must be more support for complex decision-making under the uncertainty of how to assess and manage risk to coexist with wildland fire. A long history of Canadian wildland fire research offers solid foundational knowledge related to risk, but the key knowledge gaps must be addressed to fully consider risk in a comprehensive manner. We provide a review of the current context in which risk is variably defined, and recommend use of the general paradigm where risk is the product of both the likelihood and the potential impacts of wildland fire. We then synthesize research related to wildland fire risk from the Canadian scientific literature. With this review, we aim to provide a better understanding of research challenges, limitations, and opportunities for future work on fire risk within the country.
... Lightning prediction models have employed these data to derive regression relationships with atmospheric conditions and stability indices that can be forecast with NWP. Ensemble forecasts of lightning using RF is a viable modelling approach for Alberta, Canada (Blouin, Flannigan, Wang, & Kochtubajda, 2016). Bates et al. (2017) used two machine learning methods (CART and RF) and three statistical methods to classify wet and dry thunderstorms (lightning associated with dry thunderstorms are more likely to start fires) in Australia. ...
Preprint
Full-text available
Artificial intelligence has been applied in wildfire science and management since the 1990s, with early applications including neural networks and expert systems. Since then the field has rapidly progressed congruently with the wide adoption of machine learning (ML) in the environmental sciences. Here, we present a scoping review of ML in wildfire science and management. Our objective is to improve awareness of ML among wildfire scientists and managers, as well as illustrate the challenging range of problems in wildfire science available to data scientists. We first present an overview of popular ML approaches used in wildfire science to date, and then review their use in wildfire science within six problem domains: 1) fuels characterization, fire detection, and mapping; 2) fire weather and climate change; 3) fire occurrence, susceptibility, and risk; 4) fire behavior prediction; 5) fire effects; and 6) fire management. We also discuss the advantages and limitations of various ML approaches and identify opportunities for future advances in wildfire science and management within a data science context. We identified 298 relevant publications, where the most frequently used ML methods included random forests, MaxEnt, artificial neural networks, decision trees, support vector machines, and genetic algorithms. There exists opportunities to apply more current ML methods (e.g., deep learning and agent based learning) in wildfire science. However, despite the ability of ML models to learn on their own, expertise in wildfire science is necessary to ensure realistic modelling of fire processes across multiple scales, while the complexity of some ML methods requires sophisticated knowledge for their application. Finally, we stress that the wildfire research and management community plays an active role in providing relevant, high quality data for use by practitioners of ML methods.
... Our main study area was northern Alberta (Fig. 1a). Throughout Alberta between 1961 and 2014, wildfires burned an average of 147 800 ha per year, with a minimum of 1750 ha in 1962 and maximum of 1 357 190 ha in 1981 (Blouin et al. 2016). Most of these fires occurred in the northern part of the province, which is heavily forested . ...
Article
Full-text available
Wildfires burn an average of 2 million hectares per year in Canada, most of which can be attributed to only a few days of severe fire weather. These “spread days” are often associated with large-scale weather systems. We used extreme threshold values of three Canadian Fire Weather Index System (CFWIS) variables—the fine fuel moisture code (FFMC), initial spread index (ISI), and fire weather index (FWI)—as a proxy for spread days. Then we used self-organizing maps (SOMs) to predict spread days, with sea-level pressure and 500 hPa geopotential height as predictors. SOMs require many input parameters, and we performed an experiment to optimize six key parameters. For each month of the fire season (May–August), we also tested whether SOMs performed better when trained with only one month or with neighbouring months as well. Good performance (AUC of 0.8) was achieved for FFMC and ISI, while nearly good performance was achieved for FWI. To our knowledge, this is the first study to develop a machine-learning model for extreme fire weather that could be deployed in real time.
... They can be divided into three main groups: 1) Nowcasting based on observations or on a combination of observations and Numerical Weather Prediction (NWP) model outputs. For example, Blouin et al. (2016) used a supervised machine learning method, named random forests, to forecast lightning strikes relevant to wildfire occurrence for the province of Alberta, Canada. Meng et al. (2019) developed a system that combines a set of observational data with synoptic pattern forecasting products and an electrification and discharge model. ...
Article
The study aims to assess the applicability of the current Meso-NH electrical scheme (CELLS) in the investigation of forest fire ignition. Therefore, the challenge is to diagnose cloud-to-ground (CG) lightning at 1 km spatial resolution and, subsequently, preferred regions where Forest Fire Events (FFE) could be naturally ignited. The afternoon of 17th June 2017 has been considered for the study, since the dry thunderstorm environment configured a perfect scenario for natural ignition and evolution of some fires. The Pedrógão Grande wildfire ignited in this period and was the deadliest single fire event in the Portuguese history, with at least 60 people killed. The available observational data are used to validate the numerical results and to provide a brief characterization of the meteorological environment when jointly analysed with the simulation. In addition, cloud microphysics and cloud electrical structure are explored from the model results. The spatial distribution of the simulated CG lightning showed a good agreement with the lightning strokes obtained from the national lightning detection network. Overall, this paper introduces a possible application of the Meso-NH electrical scheme, namely the study of forest fire ignition by lightning strokes.
... Over the past two decades, substantial capital and on-going operational investments were made to improve lightning detection and monitoring systems in Canada [i.e., Canadian Lightning Detection Network (CLDN)]. This facilitated the production of robust lightning climatologies (e.g., Shephard et al. 2013;Burrows and Kochtubajda 2010;Kochtubajda and Burrows 2010), implementation of new approaches to predict probabilities of future lightning occurrence (Blouin et al. 2016, Burrows et al. 2005, and the realtime communication of CG lightning threats (e.g., Mainwaring and Fricska 2016;Canadian Lightning Danger Map, ECCC 2018). ...
Article
Full-text available
Cloud-to-ground lightning is a common and dangerous natural atmospheric hazard in southern Canada. Previous research conducted by the author and colleagues, using data from 1994 to 2003, estimated that lightning directly or indirectly kills 9–10 people and injures 92–164 more each year in Canada. Repeating the analysis using data from the same government agency and media sources for the 2002–2017 period, the author found that lightning-related mortality decreased to 2–3 deaths per year, roughly 0.08 deaths per million population. An average of 180 lightning-related injuries each year (5.3 per million population) was estimated for the same period, slightly greater than the maximum documented in the 1994–2003 analysis. About half of the drop in mortality between periods may be attributed to the reduction in reported deaths associated with lightning-ignited municipal fires since 2000. The remainder may be due to a combination of greater availability and use of communication technology, faster emergency response and medical treatment, and increased public awareness of lightning hazards and safety. Further research is required to explain why lightning-related injury rates have remained stable; better understand the interaction of technological, behavioral and other factors; and to determine the efficacy of past and potential future safety interventions.
... Lightning prediction models have employed these data to derive regression relationships with atmospheric conditions and stability indices that can be forecast with NWP. Ensemble forecasts of lightning using RF is a viable modelling approach for Alberta, Canada (Blouin, Flannigan, Wang, & Kochtubajda, 2016). Bates et al. (2017) used two machine learning methods (CART and RF) and three statistical methods to classify wet and dry thunderstorms (lightning associated with dry thunderstorms are more likely to start fires) in Australia. ...
Article
Full-text available
Artificial intelligence has been applied in wildfire science and management since the 1990s, with early applications including neural networks and expert systems. Since then the field has rapidly progressed congruently with the wide adoption of machine learning (ML) methods in the environmental sciences. Here, we present a scoping review of ML applications in wildfire science and management. Our overall objective is to improve awareness of ML methods among wildfire researchers and managers, as well as illustrate the diverse and challenging range of problems in wildfire science available to ML data scientists. To that end, we first present an overview of popular ML approaches used in wildfire science to date, and then review the use of ML in wildfire science as broadly categorized into six problem domains, including: 1) fuels characterization, fire detection, and mapping; 2) fire weather and climate change; 3) fire occurrence, susceptibility, and risk; 4) fire behavior prediction; 5) fire effects; and 6) fire management. Furthermore, we discuss the advantages and limitations of various ML approaches relating to data size, computational requirements, generalizability, and interpretability, as well as identify opportunities for future advances in the science and management of wildfires within a data science context. In total, we identfied 300 relevant publications up to the end of 2019, where the most frequently used ML methods across problem domains included random forests, MaxEnt, artificial neural networks, decision trees, support vector machines, and genetic algorithms. As such, there exists opportunities to apply more current ML methods — including deep learning and agent based learning — in the wildfire sciences, especially in instances involving very large multivariate datasets. We must recognize, however, that despite the ability of ML methods to learn on their own, expertise in wildfire science is necessary to ensure realistic modelling of fire processes across multiple scales, while the complexity of some ML methods, such as deep learning, requires a dedicated and sophisticated knowledge of their application. Finally, we stress that the wildfire research and management communities play an active role in providing relevant, high quality, and freely available wildfire data for use by practitioners of ML methods.
... Subsequent investigations have analyzed lightning activity for the 1998-2000 and the 1999-2008 periods . Studies have also used CLDN data to assess lightning impacts on wildfires in the boreal regions of Yukon and the Northwest Territories (Kochtubajda et al., 2011(Kochtubajda et al., , 2019, derive relationships between convective rainfall and lightning (Kochtubajda et al., 2013), develop lightning occurrence prediction models (Blouin et al., 2016), and apply statistical techniques to create high-resolution flash density climatologies (Shephard et al., 2013). Analyses, combining the networks of the United States and Canada, have described continental-scale lightning activity (Orville et al., 2002(Orville et al., , 2011. ...
Article
Full-text available
This study presents the spatial and temporal features of more than 45 million cloud-to-ground (CG) lightning flashes recorded by the Canadian Lightning Detection Network for the years 1999–2018. Although sensor upgrades have improved the detection efficiency and location accuracy of CG lightning, the large-scale spatial patterns remain about the same as found in a previous study covering the years 1999–2008. Analyses, using equal-area squares with 10 km sides, describe the regional and seasonal characteristics of negative and positive flashes, the percentage and flash density of positive lightning, and the first-stroke peak currents of both polarities. Lightning activity over the provinces and territories is greatest in the summer, varying from 95.9% to 76.8% of the annual activity in the Northwest Territories and Ontario, respectively. Winter lightning is rare, usually occurring in extreme southern Ontario and the Atlantic Provinces, as well as over offshore regions west of Vancouver Island and the coastal waters off Nova Scotia. Preliminary analysis suggests that, compared with the 1999–2008 period, the majority of western and northern Canada has experienced more lightning days during the 2009–2018 period, whereas much of eastern Canada has experienced fewer lightning days. A statistical analysis performed on 154 stations across Canada found that the decadal increases (decreases) at 5 (31) stations were significant at the 90% confidence level or higher, and 4 (16) of these were significant at the 95% confidence level.
... Although human activity is widely recognized as the primary cause of wildfires, in most ecosystems (Amatulli et al., 2007;Archibald et al., 2009;Ganteaume et al., 2013), lightning-caused ignitions are not uncommon and the majority of burned areas in remote regions of the Boreal and Arctic biomes have a lightning origin (Krider et al., 1980;Granström, 1993). Lightning is a major cause of wildland fires in Canada (Hanes et al., 2019), and in Alberta, for example, it is responsible for igniting 45% of reported wildfires and 71% of area burned (Blouin et al., 2016). There is a clear connection between fuel dryness, its availability and the probability of a lightning ignition (Flannigan and Haar, 1986), even if complex nonenvironmental variables can play a role (Anderson, 2002). ...
Article
Full-text available
Lightning‐caused wildfires are a significant contributor to burned areas, with lightning ignitions remaining one of the most unpredictable aspects of the fire environment. There is a clear connection between fuel moisture and the probability of ignition; however, the mechanisms are poorly understood and predictive methods are underdeveloped. Establishing a lightning–ignition relationship would be useful in developing a model that would complement early warning systems designed for fire control and prevention. A machine learning (ML) approach was used to define a predictive model for wildfire ignition based on lightning forecasts and environmental conditions. Three different binary classifiers were adopted: a decision tree, an AdaBoost and a Random Forest, showing promising results, with both ensemble methods (Random Forest and AdaBoost) exhibiting an out‐of‐sample accuracy of 78%. Data provided by a Western Australia wildfire database allowed a comprehensive verification on over 145 lightning‐ignited wildfires in regions of Australia during 2016. This highlighted that in a minimum of 71% of the cases the ML models correctly predicted the occurrence of an ignition when a fire was actually initiated. The super‐learner developed is planned to be used in an operational context to the enhance information connected to fire management. A machine learning (ML) approach was used to define a predictive model for wildfire ignition based on lightning forecasts and environmental conditions. Three different binary classifiers were adopted: a decision tree, a Random Forest and an AdaBoost, showing promising results, with both ensemble methods (Random Forest and AdaBoost) exhibiting an out‐of‐sample accuracy of 78%.
... Although individual weather elements directly influence ignition potential, largescale atmospheric features and processes are also associated with lightning occurrence [13]. Lightning activity is formed due to development of atmospheric instability and to thermodynamic processes which influence its frequency and intensity [14,15]. ...
Article
Full-text available
Lightning strikes are pervasive, however, their distributions vary both spatially and in time, resulting in a complex pattern of lightning-ignited wildfires. Over the last decades, lightning-ignited wildfires have become an increasing threat in south-east Australia. Lightning in combination with drought conditions preceding the fire season can increase probability of sustained ignitions. In this study, we investigate spatial and seasonal patterns in cloud-to-ground lightning strikes in the island state of Tasmania using data from the Global Position and Tracking System (GPATS) for the period January 2011 to June 2019. The annual number of lightning strikes and the ratio of negative to positive lightning (78:22 overall) were considerably different from one year to the next. There was an average of 80 lightning days per year, however, 50% of lightning strikes were concentrated over just four days. Most lightning strikes were observed in the west and north of the state consistent with topography and wind patterns. We searched the whole population of lightning strikes for those most likely to cause wildfires up to 72 h before fire detection and within 10 km of the ignition point derived from in situ fire ignition records. Only 70% of lightning ignitions were matched up with lightning records. The lightning ignition efficiency per stroke/flash was also estimated, showing an annual average efficiency of 0.24% ignition per lightning stroke with a seasonal maximum during summer. The lightning ignition efficiency as a function of different fuel types also highlighted the role of buttongrass moorland (0.39%) in wildfire incidents across Tasmania. Understanding lightning climatology provides vital information about lightning characteristics that influence the probability that an individual stroke causes ignition over a particular landscape. This research provides fire agencies with valuable information to minimize the potential impacts of lightning-induced wildfires through early detection and effective response.
... traditionally been the tools most used for predicting lightning activity (Venzke, 2001;Mazany et al., 2002;Burrows et al., 2005;Fuelberg, 2006, 2008;Coning et al., 2011;Kuk et al., 2012;Blouin, 2014;Zepka et al., 2014). The second crucial factor affecting lightning activity is the ability of storms to generate areas with opposite electric charges via microphysical and dynamic processes, which involve hydrometeors in different stages and of varying size. ...
Article
This study presents the characteristics of cloud-to-ground lightning in the province of León (Spain), based on data collected via the lightning detection network of the Spanish Meteorological Agency. A total of 146 081 flashes and 279 220 strokes were recorded between 2000 and 2010. Spatial analysis (total, negative and positive flash density, and mean peak currents of positive and negative flashes) was performed at a resolution of 1 km. The maximum density recorded for total negative and positive flashes was 2.0 flashes km⁻² year⁻¹; 2.3 for negative flashes only and 0.178 for positive flashes. There was a different spatial distribution for positive compared with negative flashes, resulting from meteorological mechanisms involved with their polarities. The density distribution corresponding to both total and negative flashes appears to be clearly associated with topography. Interestingly, there is a clear inverse spatial correlation between the density and peak current parameters, which has important implications for constructing risk maps of lightning activity. This correlation has been quantified and confirmed for both positive and negative flashes by two separate regression equations. In the second part of the present study, a statistical model was constructed to predict lightning in the province of León, using a quadratic discriminant function that encompasses three meteorological variables obtained from National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis: lifted index, K-index and precipitable water. To construct the model, data were used from May to September over 2002–2007, and then applied to an independent sample of years from 2008 to 2010. Results were verified using skill scores probability of detection, false alarm rate, critical success index and true skill statistic. Scores obtained for the samples were 0.79, 0.45, 0.48 and 0.53 (respectively) for model construction, and 0.78, 0.14, 0.69 and 0.65 (respectively) for application to the independent sample.
... Still lightning-ignited fires are not uncommon and the majority of burned areas in remote regions of Boreal and arctic biomes have this as origin (Granström, 1993;Krider et al., 1980). Lightning is a major cause of wildland fires in Canada (Hanes et al., 2019) and in Alberta, for example, it is responsible for igniting 45% of reported wildfires and 71% of area burned (Blouin et al., 2016). Similar statistics are reported in Australia (Dowdy & Mills, 2012). ...
Article
Full-text available
Plain Language Summary Human activity is recognized as the primary cause of wildfires ignition. Still, lightning‐ignited fires are responsible for the majority of burned areas in remote regions. Unlike human behaviors, lightning activity can be predicted with a reasonable level of confidence as it is linked to weather conditions well represented in current forecasting models. Lighting predictions and environmental factors have been combined in machine‐learning based models to provide a quantitative measure to identify those episodes that are potentially conducive of fires. By providing the forecast in terms of probability of ignition rather than a binary (yes/no) outcome can highly increase the skill of the prediction. Still, the skill of a forecasting system not always equal its value. A skillful forecast can be useless for decision making if is not providing the right information, so it is important to verify that the increased skill brings real benefits. Using a simple cost‐loss model of economic value we found that for very low cost‐loss ratio (i.e., if we assume very high loss associated to the ignition) the use of probabilistic information would be also economically convenient to the decision‐maker ensuring almost 40% of the savings which would be obtained with perfect knowledge of future events.
... In the former category, recently published research includes the studies of Mostajabi et al. (2019), Shrestha et al. (2021) and Leinonen et al. (2022b). Among studies on grid-based ML prediction of lightning, Lin et al. (2019), Zhou et al. (2020), and Geng et al. (2021) used deep-learning techniques using convolutional neural networks, while Blouin et al. (2016) and La Fata et al. 2 (2021) used methods based on decision trees. Mostajabi et al. (2019) considered the nowcasting of lightning at weather station locations using ML. ...
Preprint
Full-text available
A deep learning model is presented to nowcast the occurrence of lightning at a five-minute time resolution 60 minutes into the future. The model is based on a recurrent-convolutional architecture that allows it to recognize and predict the spatiotemporal development of convection, including the motion, growth and decay of thunderstorm cells. The predictions are performed on a stationary grid, without the use of storm object detection and tracking. The input data, collected from an area in and surrounding Switzerland, comprise ground-based radar data, visible/infrared satellite data and derived cloud products, lightning detection, numerical weather prediction and digital elevation model data. We analyze different alternative loss functions, class weighting strategies and model features, providing guidelines for future studies to select loss functions optimally and to properly calibrate the probabilistic predictions of their model. Based on these analyses, we use focal loss in this study, but conclude that it only provides a small benefit over cross entropy, which is a viable option if recalibration of the model is not practical.
... The success of public messaging initiatives (e.g., fire prevention) beginning in the early 20th-century may inpart explain the progressive decline in importance of human ignitions. Compared to human ignitions, which are concentrated close to infrastructure, travel routes, and resources (Balch et al. 2017), lightning ignitions occur as clusters, often in remote locations where suppression can be logistically difficult (Wierzchowski et al. 2002;Podur and Wotton 2010;Blouin et al. 2016). In severe wildfire seasons when suppression resources may become limited, suppression of lightning-caused fires in the backcountry is often deferred in favor of suppressing those that threaten human communities and WUI in the front country (Podur and Wotton 2010). ...
Article
Full-text available
Context In fire-excluded forests across western North America, recent intense wildfire seasons starkly contrast with fire regimes of the past. The last 100 years mark a transition between pre-colonial and modern era fire regimes, providing crucial context for understanding future wildfire behavior. Objectives Using the greatest time depth of digitized fire events in Canada, we identify distinct phases of wildfire regimes from 1919 to 2019 by evaluating changes in mapped fire perimeters (> 20-ha) across the East Kootenay region (including the southern Rocky Mountain Trench), British Columbia. Methods We detect transitions in annual number of fires, burned area, and fire size; explore the role of lightning- and human-caused fires in driving these transitions; and quantify departures from historical fire frequency at the regional level. Results Relative to historical fire frequency, fire exclusion has created a significant fire deficit in active fire regimes, with a minimum of 1–10 fires missed across 46.4-percent of the landscape. Fire was active from 1919 to 1939 with frequent and large fire events, but the regime was already altered by a century of colonization. Fire activity decreased in 1940, coinciding with effective fire suppression influenced by a mild climatic period. In 2003, the combined effects of fire exclusion and accelerated climate change fueled a shift in fire regimes of various forest types, with increases in area burned and mean fire size driven by lightning. Conclusions The extent of fire regime disruption warrants significant management and policy attention to alter the current trajectory and facilitate better co-existence with wildfire throughout this century.Graphical abstract
... The fire weather predictors were sampled only with the data points that corresponded to the fire season. Before fitting a classification model, we calculated the Spearman correlation coefficient between every pair of covariates using the rcorr of the package Hmisc in R (Table S1.3) excluding those with a correlation coefficient ≥0.7 that made less ecological meaning, or that were highly correlated with more than one variable (Blouin et al., 2016;Table S1.3). The final data set consisted in 6 static variables (SLO, MAT, MAP, VEG, DNR, DNS) and 16 fire weather variables (Precipitation and temperature of every season of the previous and current year) ( Table 1). ...
Article
Full-text available
Warming trends are altering fire regimes globally, potentially impacting on the long-term persistence of some ecosystems. However, we still lack clear understanding of how climatic stressors will alter fire regimes along productivity gradients. We trained a Random Forests model of fire probabilities across a 5°lat × 2° long trans-Andean rainfall gradient in northern Patagonia using a 23-year long fire record and biophysical, vegetation, human activity and seasonal fire weather predictors. The final model was projected onto mid- and late 21st century fire weather conditions predicted by an ensemble of GCMs using 4 emission scenarios. We finally assessed the vulnerability of different forest ecosystems by matching predicted fire return intervals with critical forest persistence fire return thresholds developed with landscape simulations. Modern fire activity showed the typical hump-shaped relationship with productivity and a negative distance relationship with human settlements. However, fire probabilities were far more sensitive to current season fire weather than to any other predictor. Sharp responsiveness of fire to the accelerating drier/warmer fire seasons predicted for the remainder of the 21st century in the region led to 2 to 3-fold (RCPs 4.5 and 8.5) and 3 to 8-fold increases in fire probabilities for the mid- and late 21st century, respectively. Contrary to current generalizations of larger impacts of warming on fire activity in fuel-rich ecosystems, our modeling results showed first an increase in predicted fire activity in less productive ecosystems (shrublands and steppes) and a later evenly amplified fire activity-productivity relationship with it shape resembling (at higher fire probabilities) the modern hump-shaped relationship. Despite this apparent homogeneous effect of warming on fire activity, vulnerability to predicted late 21st century shorter fire intervals were higher in most productive ecosystems (subalpine deciduous and evergreen Nothofagus-dominated rainforests) due to a general lack of fire-adapted traits in the dominant trees that compose these forests.
... The fire weather predictors were sampled only with the data points that corresponded to the fire season. Before fitting a classification model, we calculated the Spearman correlation coefficient between every pair of covariates using the rcorr of the package Hmisc in R (Table S1.3) excluding those with a correlation coefficient ≥0.7 that made less ecological meaning, or that were highly correlated with more than one variable (Blouin et al., 2016;Table S1.3). The final data set consisted in 6 static variables (SLO, MAT, MAP, VEG, DNR, DNS) and 16 fire weather variables (Precipitation and temperature of every season of the previous and current year) ( Table 1). ...
... Throughout boreal Canada, anthropogenic factors increase fire probabil-ity , with humans igniting most fires close to roads while lightning-caused fires are responsible for the majority of burned area in the more remote locations (Gralewicz et al., 2012). Blouin et al. (2016) found that 45 % of wildfires in Alberta were started by lightning but were responsible for 71 % of burned area. In Finland, lightning-caused fires account for less than 15 % of forest fires (Larjavaara et al., 2005). ...
Article
Full-text available
In recent years, the pan-Arctic region has experienced increasingly extreme fire seasons. Fires in the northern high latitudes are driven by current and future climate change, lightning, fuel conditions, and human activity. In this context, conceptualizing and parameterizing current and future Arctic fire regimes will be important for fire and land management as well as understanding current and predicting future fire emissions. The objectives of this review were driven by policy questions identified by the Arctic Monitoring and Assessment Programme (AMAP) Working Group and posed to its Expert Group on Short-Lived Climate Forcers. This review synthesizes current understanding of the changing Arctic and boreal fire regimes, particularly as fire activity and its response to future climate change in the pan-Arctic have consequences for Arctic Council states aiming to mitigate and adapt to climate change in the north. The conclusions from our synthesis are the following. (1) Current and future Arctic fires, and the adjacent boreal region, are driven by natural (i.e. lightning) and human-caused ignition sources, including fires caused by timber and energy extraction, prescribed burning for landscape management, and tourism activities. Little is published in the scientific literature about cultural burning by Indigenous populations across the pan-Arctic, and questions remain on the source of ignitions above 70∘ N in Arctic Russia. (2) Climate change is expected to make Arctic fires more likely by increasing the likelihood of extreme fire weather, increased lightning activity, and drier vegetative and ground fuel conditions. (3) To some extent, shifting agricultural land use and forest transitions from forest–steppe to steppe, tundra to taiga, and coniferous to deciduous in a warmer climate may increase and decrease open biomass burning, depending on land use in addition to climate-driven biome shifts. However, at the country and landscape scales, these relationships are not well established. (4) Current black carbon and PM2.5 emissions from wildfires above 50 and 65∘ N are larger than emissions from the anthropogenic sectors of residential combustion, transportation, and flaring. Wildfire emissions have increased from 2010 to 2020, particularly above 60∘ N, with 56 % of black carbon emissions above 65∘ N in 2020 attributed to open biomass burning – indicating how extreme the 2020 wildfire season was and how severe future Arctic wildfire seasons can potentially be. (5) What works in the boreal zones to prevent and fight wildfires may not work in the Arctic. Fire management will need to adapt to a changing climate, economic development, the Indigenous and local communities, and fragile northern ecosystems, including permafrost and peatlands. (6) Factors contributing to the uncertainty of predicting and quantifying future Arctic fire regimes include underestimation of Arctic fires by satellite systems, lack of agreement between Earth observations and official statistics, and still needed refinements of location, conditions, and previous fire return intervals on peat and permafrost landscapes. This review highlights that much research is needed in order to understand the local and regional impacts of the changing Arctic fire regime on emissions and the global climate, ecosystems, and pan-Arctic communities.
... In order to improve the prediction power of these models, additional bioclimate variables were added to the analysis to expand on the climatic differences within the ecozone delineations. The majority of the initial 36 bioclimatic variables were highly correlated with each other based on Pearson's correlation coefficient (r $ 0.7; Dormann et al. 2013;Blouin et al. 2016). As a result, only nine bioclimatic variables were selected for the final analysis where r , 0.7 (Table 3). ...
Article
In Canada, fire behaviour is modelled based on a fuel classification system of 16 fuel types. Average fuel loads are used to represent a wide range of variability within each fuel type, which can lead to inaccurate predictions of fire behaviour. Dead and down woody debris (DWD) is a major component of surface fuels affecting surface fuel consumption, potential crown fire initiation, and resulting crown fuel consumption and overall head fire intensity. This study compiled a national database of DWD fuel loads and analysed it for predictive driving variables. The database included DWD fuel loads for all dominant Canadian forest types at three size classes: fine (<1 cm), medium (1–7 cm) and coarse (>7 cm). Predictive models for DWD fuel load by size classes individually and collectively for various forest types and ecozones were analysed. Bioclimatic regime, age, spatial position, drainage, and structural components including diameter at breast height and stem density were significant variables. This study provides tools to improve our understanding of the spatial distribution of DWD across Canada, which will enhance our ability to represent its contribution within fire behaviour and fire effects models.
Preprint
We introduce a class of proper scoring rules for evaluating spatial point process forecasts based on summary statistics. These scoring rules rely on Monte-Carlo approximations of expectations and can therefore easily be evaluated for any point process model that can be simulated. In this regard, they are more flexible than the commonly used logarithmic score and other existing proper scores for point process predictions. The scoring rules allow for evaluating the calibration of a model to specific aspects of a point process, such as its spatial distribution or tendency towards clustering. Using simulations we analyze the sensitivity of our scoring rules to different aspects of the forecasts and compare it to the logarithmic score. Applications to earthquake occurrences in northern California, USA and the spatial distribution of Pacific silver firs in Findley Lake Reserve in Washington, USA highlight the usefulness of our scores for scientific model selection.
Conference Paper
This paper discusses the use of Random Forest (RF), a popular Machine Learning (ML) algorithm, to perform spatially explicit nowcasting of cloud-to-ground lightning occurrence. An application to the Italian territory and the surrounding seas is then presented. Specifically, a dataset including eighteen geo-environmental features has been used to forecast 1-hour ahead lightning occurrence over a three-months period (August- October 2018). The feature importance resulting from the best RF model showed how data-driven models are able to identify relationships between variables in agreement with previous physically-based knowledge of the phenomenon. The encouraging results obtained in terms of forecasting Accuracy suggest how, after proper improvements, ML-based algorithms could find their place in wider earlywarning systems to support disaster risk management procedures.
Preprint
Full-text available
In recent years, the Pan-Arctic region has experienced increasingly extreme fire seasons. Fires in the northern high latitudes are driven by current and future climate change, lightning, fuel conditions, and human activity. In this context, conceptualizing and parameterizing current and future Arctic fire regimes will be important for fire and land management as well as understanding current and predicting future fire emissions. The objectives of this review were driven by policy questions identified by the Arctic Monitoring and Assessment Programme (AMAP) Working Group and posed to its Expert Group on Short-Lived Climate Forcers. This review synthesises current understanding of the changing Arctic and boreal fire regimes, particularly as fire activity and its response to future climate change in the Pan-Arctic has consequences for Arctic Council states aiming to mitigate and adapt to climate change in the north. The conclusions from our synthesis are the following: (1) Current and future Arctic fires, and the adjacent boreal region, are driven by natural (i.e., lightning) and human-caused ignition sources, including fires caused by timber and energy extraction, prescribed burning for landscape management, and tourism activities. Little is published in the scientific literature about cultural burning by Indigenous populations across the Pan-Arctic and questions remain on the source of ignitions above 70° N in Arctic Russia. (2) Climate change is expected to make Arctic fires more likely by increasing the likelihood of extreme fire weather, increased lightning activity, and drier vegetative and ground fuel conditions. (3) To some extent, shifting agricultural land use, forest-steppe to steppe, tundra-to-taiga, and coniferous-to-deciduous forest transitions in a warmer climate may increase and decrease open biomass burning. However, at the country- and landscape-scales, these relationships are not well established. (4) Current black carbon and PM2.5 emissions from wildfires above 50° N and 65° N are larger than emissions from the anthropogenic sectors of residential combustion, transportation, and flaring, respectively. Wildfire emissions have increased from 2010 to 2020, particularly above 60° N, with 56 % of black carbon emissions above 65° N in 2020 attributed to open biomass burning – indicating how extreme the 2020 wildfire season was and future Arctic wildfire seasons potential. (5) What works in the boreal zones to prevent and fight wildfires may not work in the Arctic. Fire management will need to adapt to a changing climate, economic development, the Indigenous and local communities, and fragile northern ecosystems, including permafrost and peatlands. (6) Factors contributing to the uncertainty of predicting and quantifying future Arctic fire regimes include underestimation of Arctic fires by satellite systems, lack of agreement between Earth observations and official statistics, and still needed refinements of location, conditions, and previous fire return intervals on peat and permafrost landscapes. This review highlights that much research is needed in order to understand the local and regional impacts of the changing Arctic fire regime on emissions and the global climate, ecosystems and Pan-Arctic communities.
Article
Forest organic layers are important soil carbon pools that can, in the absence of disturbance, accumulate to great depths, especially in lowland areas. Across the Canadian boreal forest, fire is the primary disturbance agent, often limiting organic layer accumulation through the direct consumption of these fuels. Organic layer thickness (OLT) and fuel load (OLFL) are common physical attributes used to characterize these layers, especially for wildland fire science. Understanding the drivers and spatial distribution of these attributes is important to improve predictions of fire behaviour, emissions and effects models. We developed maps of OLT and OLFL using machine learning approaches (weighted K-nearest neighbour and random forests) for the forested region of the province of Alberta, Canada (538, 058 km²). The random forests approach was found to be the best approach to model the spatial distribution of these forest floor attributes. A databased of 3, 237 OLT and 594 OLFL plots were used to train the models. The error in our final model, particularly for OLT (5 cm), was relatively close to the variability we would expect to find naturally (3 cm). The dominant tree species was the most important covariate in the models. Age, solar radiation, spatial location, climate variables and surficial geology were also important drivers, although their level of importance varied between tree species and depended on the modelling method that was used.
Article
Full-text available
Wildland fires exert substantial impacts on tundra ecosystems of the high northern latitudes (HNL), ranging from biogeochemical impact on climate system to habitat suitability for various species. Cloud-to-ground (CG) lightning is the primary ignition source of wildfires. It is critical to understand mechanisms and factors driving lightning strikes in this cold, treeless environment to support operational modeling and forecasting of fire activity. Existing studies on lightning strikes primarily focus on Alaskan and Canadian boreal forests where land-atmospheric interactions are different and, thus, not likely to represent tundra conditions. In this study, we designed an empirical-dynamical method integrating Weather Research and Forecast (WRF) simulation and machine learning algorithm to model the probability of lightning strikes across Alaskan tundra between 2001 and 2017. We recommended using Thompson 2-moment and Mellor-Yamada-Janjic schemes as microphysics and planetary boundary layer parameterizations for WRF simulations in the tundra. Our modeling and forecasting test results have shown a strong capability to predict CG lightning probability in Alaskan tundra, with the values of area under the receiver operator characteristics curves above 0.9. We found that parcel lifted index and vertical profiles of atmospheric variables, including geopotential height, dew point temperature, relative humidity, and velocity speed, important in predicting lightning occurrence, suggesting the key role of convection in lightning formation in the tundra. Our method can be applied to data-scarce regions and support future studies of fire potential in the HNL.
Article
This work presents the first observation of a multi-stroke positive cloud-to-ground lightning flash sharing the same channel to ground mapped with a very high frequency broadband interferometer and a Lightning Mapping Array. This type of lightning flash is very rarely observed, and it is currently unclear how frequent it is and even under what conditions it occurs. Our observations indicate a scenario where the first downward positive leader initiates from a decayed negative channel. After the first return stroke, some of the main negative channel branches stop propagating and are likely cut off. A fast recoil leader and/or a fast breakdown play a crucial role in reconnecting these previously decayed leader channels and initiating the subsequent positive stroke. The mechanism we propose to describe the phenomenon allows us to explain its rarity and the discrete positive charge transfer to the ground.
Preprint
Full-text available
ContextIn fire-excluded forests across western North America, recent intense wildfire seasons starkly contrast with fire regimes of the past. The last 100 years mark a transition between pre-colonial and modern era fire regimes, providing crucial context for understanding future wildfire behavior.Objectives Using the greatest time depth of digitized fire events in Canada, we identify distinct phases of wildfire regimes from 1919 to 2019 by evaluating changes in mapped fire perimeters (>20-ha) across the East Kootenay region (including the southern Rocky Mountain Trench), British Columbia. Methods We detect transitions in annual number of fires, burned area, and fire size; explore the role of lightning- and human-caused fires in driving these transitions; and quantify departures from historical fire frequency at the regional level.ResultsRelative to historical fire frequency, fire exclusion has created a significant fire deficit in active fire regimes, with a minimum of 1–10 fires missed across 46.4-percent of the landscape. Fire was active from 1919 to 1939 with frequent and large fire events, but the regime was already altered by a century of colonization. Fire activity decreased in 1940, coinciding with effective fire suppression influenced by a mild climatic period. In 2003, the combined effects of fire exclusion and accelerated climate change fueled a shift in fire regimes of various forest types, with increases in area burned and mean fire size driven by lightning.Conclusions The extent of fire regime disruption warrants significant management and policy attention to alter the current trajectory and facilitate better co-existence with wildfire throughout this century.
Article
A method to predict lightning by machine learning analysis of atmospheric electric fields is proposed for the first time. In this study, we calculated an anomaly score with long short-term memory (LSTM), a recurrent neural network analysis method, using electric field data recorded every second on the ground. The threshold value of the anomaly score was defined, and a lightning alarm at the observation point was issued or canceled. Using this method, it was confirmed that 88.9% of lightning occurred while alarming. These results suggest that a lightning prediction system with an electric field sensor and machine learning can be developed in the future.
Article
Full-text available
Forest fires destroy the ecological environment and cause large property loss. There is much research in the field of geographic information that revolves around forest fires. The traditional forest fire prediction methods hardly consider multi-source data fusion. Therefore, the forest fire predictions ignore the complex dependencies and correlations of the spatiotemporal kind that usually bring valuable information for the predictions. Although the knowledge graph methods have been used to model the forest fires data, they mainly rely on artificially defined inference rules to make predictions. There is currently a lack of a representation and reasoning methods for forest fire knowledge graphs. We propose a knowledge-graph- and representation-learning-based forest fire prediction method in this paper for addressing the issues. First, we designed a schema for the forest fire knowledge graph to fuse multi-source data, including time, space, and influencing factors. Then, we propose a method, RotateS2F, to learn vector-based knowledge graph representations of the forest fires. We finally leverage a link prediction algorithm to predict the forest fire burning area. We performed an experiment on the Montesinho Natural Park forest fire dataset, which contains 517 fires. The results show that our method reduces mean absolute deviation by 28.61% and root-mean-square error by 53.62% compared with the previous methods.
Article
Full-text available
In Canadian forests, the majority of burned area occurs on a small number of days of extreme fire weather. These days lie within the tail end of the distribution of fire weather, and are often the periods when fire suppression capacity is most challenged. We examined the historic and future frequency of such extreme fire weather events across 16 fire regime zones in the forested regions of Canada from 1970 to the year 2090. Two measurements are used to measure the extreme fire weather events, the 95th percentile of Fire Weather Index (FWI 95) and the number of spread days. The annual frequency of fire spread days is modelled to increase 35–400 % by 2050 with the greatest absolute increases occurring in the Boreal Plains of Alberta and Saskatchewan. The largest proportional increase in the number of spread days is modelled to occur in coastal and temperate forests. This large increase in spread days was found despite a modest average increase in FWI 95. Our findings suggest that the impact of future climate change in Canadian forests is sufficient to increase the number of days with active fire spread. Fire management agencies in coastal and temperate regions may need to adapt their planning and capacity to deal with proportionally larger changes to their fire weather regime compared to the already high fire management capacity found in drier continental regions.
Article
Full-text available
Comparisons are made between thunderstorm data collected from a lightning detector network and from conventional climatic stations for the province of Manitoba, Canada. The greater resolution in time and space of lightning detector (direction finder) data makes it a valuable source of thunderstorm information and lends itself to some important applications. Data were collected for the forest fire season of 1985 using a network of 7 lightning direction finders distributed throughout the province. Some 67,912 cloud-to-ground lightning strikes were recorded by time and location during 122 thunderstorm days. July was the most active month with 27,260 strikes over 28 days. Two regions of the province had the greatest concentration of lightning strikes, indicating some influence by topography and position of large lakes. Case studies are presented of the most active lightning storms of 1985 and 1986. These storms are exclusively frontal storms, with most having similar synoptic weather patterns to those of large hailstorms and tornadoes in Manitoba. Relationships between meteorological parameters and lightning strike distribution are presented. These relationships may prove useful in the suppression of lightning-caused forest fires, especially in remote areas of the province. [Key words: lightning, thunderstorm, synoptic climatology, natural hazards, fire prevention.]
Article
Full-text available
Cloud-to-ground lightning data have been analyzed for the years 1998-2000 for North America (Canada plus the contiguous United States) for all ground flashes, positive flashes, the percentage of positive lightning, peak currents for negative and positive lightning, and for negative and positive multiplicity. The authors examined a total of 88.4 million flashes divided among the three years, 30.6 million (1998), 29.6 million (1999), and 28.3 million (2000). The highest flash densities, uncorrected for flash detection efficiency, occur in the provinces of Alberta and Saskatchewan (1-3 flashes km-2) in Canada, and along the Gulf Coast and in Florida (exceeding 9 flashes km-2) in the United States. Maximum positive flash densities in Canada range from 0.1 to 0.3 flashes km-2 and in the United States, to over 0.7 flashes km-2 (areas in the Midwest, the Gulf Coast, and Florida).
Article
Full-text available
Lightning is the primary cause of fire in the forested regions of the Pacific Northwest, especially when it occurs without significant precipitation at the surface. Using thunderstorm occurrence and precipitation observations for the period 1948-77, along with automated lightning strike data for the period 1986-96, it was possible to classify convective days as either `dry' or `wet' for several stations in the Pacific Northwest. Based on the classification, a discriminant analysis was performed on coincident upper-air sounding data from Spokane, Washington. It was found that a discriminant rule using the dewpoint depression at 85 kPa and the temperature difference between 85 and 50 kPa was able to classify correctly between 56% and 80% of the convective days as dry or wet. Also, composite maps of upper-air data showed distinctly different synoptic patterns among dry days, wet days, and all days. These findings potentially can be used by resource managers to gain a greater understanding of the atmospheric conditions that are conducive to lightning-induced fires in the Pacific Northwest.
Article
Full-text available
A Large Fire Database (LFDB), which includes information on fire location, start date, final size, cause, and suppression action, has been developed for all fires larger than 200 ha in area for Canada for the 1959-1997 period. The LFDB represents only 3.1% of the total number of Canadian fires during this period, the remaining 96.9% of fires being suppressed while <200 ha in size, yet accounts for ∼97% of the total area burned, allowing a spatial and temporal analysis of recent Canadian landscape-scale fire impacts. On average ∼2 million ha burned annually in these large fires, although more than 7 million ha burned in some years. Ecozones in the boreal and taiga regions experienced the greatest areas burned, with an average of 0.7% of the forested land burning annually. Lightning fires predominate in northern Canada, accounting for 80% of the total LFDB area burned. Large fires, although small in number, contribute substantially to area burned, most particularly in the boreal and taiga regions. The Canadian fire season runs from late April through August, with most of the area burned occurring in June and July due primarily to lightning fire activity in northern Canada. Close to 50% of the area burned in Canada is the result of fires that are not actioned due to their remote location, low values-at-risk, and efforts to accommodate the natural role of fire in these ecosystems. The LFDB is updated annually and is being expanded back in time to permit a more thorough analysis of long-term trends in Canadian fire activity.
Article
Full-text available
The primary weather forecast challenge at the Cape Canaveral Air Station and Kennedy Space Center is lightning. This paper describes a statistical approach that combines integrated precipitable water vapor (IPWV) data from a global positioning system (GPS) receiver site located at the Kennedy Space Center (KSC) with other meteorological data to develop a new GPS lightning index. The goal of this effort is to increase the forecasting skill and lead time for prediction of a first strike at the KSC. Statistical regression methods are used to identify predictors that contribute skill in forecasting a lightning event. Four predictors were identified out of a field of 23 predictors that were tested, determined using data from the 1999 summer thunderstorm season. They are maximum electric field mill values, GPS IPWV, 9-h change in IPWV, andK index. The GPS lightning index is a binary logistic regression model made up of coefficients multiplying the four predictors. When time series of the GPS lightning index are plotted, a common pattern emerges several hours prior to a lightning event. Whenever the GPS lightning index falls to 0.7 or below, lightning occurs within the next 12.5 h. An index threshold value of 0.7 was determined from the data for lightning prediction. Forecasting time constraints based on KSC weather notification requirements were incorporated into the verification. Forecast verification results obtained by using a contingency table revealed a 26.2% decrease from the KSC's previous-season false alarm rates for a nonindependent period and a 13.2% decrease in false alarm rates for an independent test season using the GPS lightning index. In addition, the index improved the KSC desired lead time by nearly 10%. Although the lightning strike window of 12 h is long, the GPS lightning index provides useful guidance to the forecaster in preparing lighting forecasts, when combined with other resources such as radar and satellite data. Future testing of the GPS lightning index and the prospect of using the logistic regression approach in forecasting related weather hazards are discussed.
Article
Full-text available
We summarize the temporal and spatial characteristics of polarity, multiplicity and first‐stroke peak current of approximately 23.5 million cloud‐to‐ground (CG) lightning flashes detected by the Canadian Lightning Detection Network for the period 1999–2008. Regional differences in these parameters reflect the complex nature and structure of thunderstorms across the country. The annual mean percentage of positive CG flashes was found to be lowest in eastern Canada (11%) and highest in northern Canada (17%). The data do not show any trends over the years in any region. The monthly distribution of positive CG flashes reflects a strong seasonality in all regions, with higher values in winter than in summer. Areas of more than 25% positive flashes are observed along the west coast of British Columbia, in Yukon extending southeast into central British Columbia, in southern Manitoba, northern Quebec, Newfoundland and off the coast of Nova Scotia. The percentages of single‐stroke positive and negative flashes for northern (western, eastern) Canada are 93% and 63%, (89% and 48%, 90% and 50%), respectively. The monthly distribution of multiplicity for negative CG flashes peaks between 2 and 2.4 strokes per flash in the summer and early fall in all regions. The multiplicity of positive flashes (slightly higher than 1 stroke per flash) shows little variation throughout the year in all regions. The annual variation of median negative and positive first‐stroke peak currents reflects a latitudinal dependence over the past decade. The lowest values for each polarity are observed in southern Canada and the highest values occur in the North. The data do not show any trends in peak currents over the years in the eastern or western regions of Canada. The monthly median first‐stroke peak currents for both polarities are strongest in winter and reach a minimum during summer in all regions. Large current flashes ≥100 kA are usually detected in summer and comprise less than 1% of the average annual CG flashes detected in Canada. Large current flashes with stroke multiplicity ≥10 are usually associated with negative polarity. These CG flashes are mostly detected in western Canada.
Article
Full-text available
This paper presents a model to predict the probability that a lightning flash will lead to a detectable fire. This is done by estimating the probability of the lightning flash having a long-continuing current, the probability of ignition, the probability of survival, and the probability of arrival. Individual probabilities are calculated using the lightning, noon weather, and forest inventory data and combined to predict the number of ignitions, holdovers, and detectable fires within a region. The model was run for six fire seasons in Saskatchewan and predicted results were compared with the actual number of fires for that season. The model successfully predicted the number of fires on 55.7% of the days with a 64.8% detection rate and a false alarm rate of 29.7%. The model was found to be highly sensitive to moisture conditions, resulting in some unusually high predictions under dry conditions.
Article
Full-text available
Statistical models valid May–September were developed to predict the probability of lightning in 3-h intervals using observations from the North American Lightning Detection Network and predictors derived from Global Environmental Multiscale (GEM) model output at the Canadian Meteorological Centre. Models were built with pooled data from the years 2000–01 using tree-structured regression. Error reduction by most models was about 0.4–0.7 of initial predictand variance. Many predictors were required to model lightning occurrence for this large area. Highest ranked overall were the Showalter index, mean sea level pressure, and troposphere precipitable water. Three-hour changes of 500-hPa geopotential height, 500–1000-hPa thickness, and MSL pressure were highly ranked in most areas. The 3-h average of most predictors was more important than the mean or maximum (minimum where appropriate). Several predictors outranked CAPE, indicating it must appear with other predictors for successful statistical lightning prediction models. Results presented herein demonstrate that tree-structured regression is a viable method for building statistical models to forecast lightning probability. Real-time forecasts in 3-h intervals to 45–48 h were made in 2003 and 2004. The 2003 verification suggests a hybrid forecast based on a mixture of maximum and mean forecast probabilities in a radius around a grid point and on monthly climatology will improve accuracy. The 2004 verification shows that the hybrid forecasts had positive skill with respect to a reference forecast and performed better than forecasts defined by either the mean or maximum probability at most times. This was achieved even though an increase of resolution and change of convective parameterization scheme were made to the GEM model in May 2004.
Article
Full-text available
Since February 1998 the Canadian Lightning Detection Network has provided continuous lightning detection in space and time to about 65°N in the west, 55°N in the east, and offshore to about 300 km. From February 1999 the distinction between cloud‐to‐ground and cloud‐to‐cloud lightning discharges was added. Approximately 21 million lightning flashes from February 1998 to December 2000 were sorted by equal‐area square grid cells with 20‐km sides and analyzed for Canada and the adjacent United States. Both cloud‐to‐ground lightning and cloud‐to‐cloud lightning were studied, as well as total activity irrespective of lightning discharge type. We derived mean patterns of number of lightning occurrence days, flash density, fraction of cloud‐to‐cloud lightning, fraction of nocturnal‐morning lightning, duration, ratio of positive to negative cloud‐to‐ground flashes, peak current, and mean multiplicity of cloud‐to‐ground flashes. A complex national pattern of lightning occurrence was revealed, showing strong regional, diurnal and seasonal dependencies. There are substantial influences by elevated terrain features and major land‐water boundaries. Three regions where significant features were seen in lightning patterns were studied in greater detail: the Alberta foothills to western Saskatchewan, southern Ontario and vicinity, and south‐east of Nova Scotia. For these areas there were several features observed in all three years or in two of the three years. For those fields which we could compare with results of previous studies for the northern United States, our conclusions were similar even though different years were analyzed. While definition of a true Canadian lightning climatology is likely several years away, we believe we have identified many of its aspects, and where possible we suggest physical reasons for significant features seen.
Article
Full-text available
This study investigates the relationship between lightning activity and the occurrence of lightning-ignited forest fires in the Northwestern Region of Ontario. We found that the Duff Moisture Code (a component of the Fire Weather Index System) and the multiplicity of the negative lightning discharges were the most important variables for estimating the number of lightning-ignited fires on a daily basis for Universal Transverse Mercator zone 15 in Ontario. Also, the results indicate that negative lightning ignited more fires than positive lightning discharges, which is contrary to popular belief. Nearly 50% of the variance in the forest fire occurrence data was explained using linear stepwise regression. Future work will focus on finer temporal and spatial scales.
Article
Full-text available
This study examines the influences of fuel, weather and topography on lightning-caused forest fires in portions of southern British Columbia and Alberta, Canada. The results show a significant difference in lightning and lightning-caused fires east and west of the Continental Divide. In British Columbia, on average there was one fire for every 50 lightning discharges whereas in Alberta there was one fire for every 1400 lightning discharges. Elevation, the distribution of lightning strikes, the Daily Severity Rating (a component of the Canadian Forest Fire Weather Index System) and vegetation composition were identified as primary agents controlling lightning fire occurrence. However, the multivariate analysis does suggest that there are other factors influencing fire occurrence other than the biophysical factors we tested. The implications of the lightning and lightning-ignited fires for land managers are discussed.
Article
Full-text available
Wildland fire is a global phenomenon, and a result of interactions between climate–weather, fuels and people. Our climate is changing,rapidly primarily through the release of greenhouse,gases that may,have profound,and possibly unexpected,impacts on global fire activity. The present paper reviews the current understanding,of what the future may bring with respect to wildland fire and discusses future options for research and management. To date, research suggests a general increase in area burned and fire occurrence but there is a lot of spatial variability, with some areas of no change or even decreases in area burned,and occurrence. Fire seasons are lengthening,for temperate,and boreal regions and this trend should continue in a warmer,world. Future trends of fire severity and intensity are difficult to determine,owing,to the complex and non-linear interactions between weather, vegetation and people. Improved fire data are required along with continued global studies that dynamically include weather, vegetation, people, and other disturbances. Lastly, we need more research on the role of policy, practices and human behaviour because most of the global fire activity is directly attributable to people. Additional keywords: area burned, carbon, emissions, fire activity, forest fire, intensity, management, modelling,
Article
Full-text available
Wildfire impacts on ecological and socioeconomic systems are regulated, in part, by climate. Association between hemispheric-scale climate patterns and annual wildfire activity can be obscured by local factors that also control the initiation and spread of fires. Vegetation, topography, and fire suppression can be expected to influence conventional measures of annual wildfire activity such as area burned, effectively concealing evidence of broad-scale climate influences. This study investigates alternatives to area-burned statistics for quantifying annual wildfire activity in Canada in relation to Northern Hemisphere climate variability represented by the Atlantic Multidecadal Oscillation (AMO). We depart from conventional approaches by including socioeconomic measures of wildfire activity and by assessing spatially referenced wildfire data over units of observation chosen explicitly to diminish variability caused by factors unrelated to broad-scale climate. Our data-centred approach, combined with linear regression modelling, revealed that the AMO was positively correlated with national time series of very large fires (≥10 000 ha), wildfire-related evacuations, and fire suppression expenditures over the period 1975–2007. The AMO and wildfire activity were most closely coupled during a period of predominantly positive-phase Arctic Oscillation (AO) and Pacific Decadal Oscillation (PDO) between 1989 and 2001. Positive correlation between maximum evacuation wind speed and the AMO suggests that wind may be a causal factor in the AMO–wildfire relationship.
Article
Full-text available
The NCEP-DOE Atmospheric Model Intercomparison Project (AMIP-II) reanalysis is a follow-on project to the "50-year" (1948-present) NCEP-NCAR Reanalysis Project. NCEP-DOE AMIP-II reanalysis covers the "20-year" satellite period of 1979 to the present and uses an updated forecast model, updated data assimilation system, improved diagnostic outputs, and fixes for the known processing problems of the NCEP-NCAR reanalysis. Only minor differences are found in the primary analysis variables such as free atmospheric geopotential height and winds in the Northern Hemisphere extratropics, while significant improvements upon NCEP-NCAR reanalysis are made in land surface parameters and land-ocean fluxes. This analysis can be used as a supplement to the NCEP-NCAR reanalysis especially where the original analysis has problems. The differences between the two analyses also provide a measure of uncertainty in current analyses.
Article
Full-text available
We hâve mode a preliminary study of cloud-to-grou nd lightning over southern Ontario and thé adjoining Gréât Lakes région. The lightning data set, using magnetic di- rection finding, is sufficiently accurate to study lightning climatology. Cloud-to-ground flash totals hâve been found for thé three warm seasons 1989-91. A large variation in flash total, lightning-day frequency and number ofhigh flash density storms occurs over thé area, with thé maximum in southwestern Ontario. The area of thé maximum also has a strong diurnal cycle and relatively few positive flashes. Several physical causes may contribute to this. Lake areas usually hâve slightfy fewer flashes than nearby land areas and warm water usually has more flashes than cold water. The Gréât Lakes do produce more lightning than océan areas. Convergence Unes of lake breezes and other lake circulations can, however, be sites for storms with intense lightning. High surface température and moisture leads to an increase in lightning génération. Over land, upslope flow increases lightning-producing storms and downslope flow decreases them. High flash density storms may be favoured by smooth rather than rough ground, and by open farmland rather than forest. On thé other hand, there does not seem to be a clear urban effect increasing lightning in thé Gréât Lakes
Article
Full-text available
Breiman (2001a,b) has recently developed an ensemble classification and regression approach that displayed outstanding performance with regard prediction error on a suite of benchmark datasets. As the base constituents of the ensemble are tree-structured predictors, and since each of these is constructed using an injection of randomness, the method is called 'random forests'. That the exceptional performance is attained with seemingly only a single tuning parameter, to which sensitivity is minimal, makes the methodology all the more remarkable. The individual trees comprising the forest are all grown to maximal depth. While this helps with regard bias, there is the familiar tradeoff with variance. However, these variability concerns were potentially obscured because of an interesting feature of those benchmarking datasets extracted from the UCI machine learning repository for testing: all these datasets are hard to overfit using tree-structured methods. This raises issues about the scope of the repository. With this as motivation, and coupled with experience from boosting methods, we revisit the formulation of random forests and investigate prediction performance on real-world and simulated datasets for which maximally sized trees do overfit. These explorations reveal that gains can be realized by additional tuning to regulate tree size via limiting the number of splits and/or the size of nodes for which splitting is allowed. Nonetheless, even in these settings, good performance for random forests can be attained by using larger (than default) primary tuning parameter values. Abstract. Breiman (2001a,b) has recently developed an ensemble classification and regression approach that displayed outstanding performance with regard prediction error on a suite of benchmark datasets. As the base constituents of the ensemble are tree-structured predictors, and since each of these is constructed using an injection of randomness, the method is called 'random forests'. That the exceptional performance is attained with seemingly only a single tuning pa-rameter, to which sensitivity is minimal, makes the methodology all the more remarkable. The individual trees comprising the forest are all grown to maximal depth. While this helps with regard bias, there is the familiar tradeoff with variance. However, these variability concerns were potentially obscured because of an interesting feature of those benchmarking datasets extracted from the UCI machine learning repository for testing: all these datasets are hard to overfit using tree-structured methods. This raises issues about the scope of the repository. With this as motivation, and coupled with experience from boosting methods, we revisit the formulation of random forests and investigate prediction performance on real-world and simu-lated datasets for which maximally sized trees do overfit. These explorations reveal that gains can be realized by additional tuning to regulate tree size via limiting the number of splits and/or the size of nodes for which splitting is allowed. Nonetheless, even in these settings, good performance for random forests can be attained by using larger (than default) primary tuning parameter values.
Article
Full-text available
This article summarizes research completed to assess the risk of lightning-related injuries and fatalities in Canada. Although lightning mortality has declined significantly over the past century, it remains a common meteorological hazard that regularly kills and injures. Based on an analysis of media reports, vital statistics, hospital admission and emergency room visit records, and fire loss data, the authors estimate that on average about 9–10 lightning-related deaths and 92–164 injuries occur each year in Canada. The distribution of casualties reflects current provincial population and cloud-to-ground lightning densities. Consistent with similar studies in other developed nations, most lightning-related fatalities and injuries in Canada occur during the June-August summer season, coincident with peak lightning, and during the Thursday-Saturday period, most likely related to higher rates of participation in outdoor activities. The majority of victims are male, less than 46years old, and engaged in outdoor recreational activities when injured or killed in a lightning incident. Media reports used in the study were found to underestimate both lightning mortality (36%) and morbidity (20–600%).
Article
Full-text available
Submitted to the Faculty of Graduate Studies and Research in partial fulfilment of the requirements for the degree of Master of Science, Department of Geography. Thesis (M.Sc.)--University of Alberta, 1991.
Article
For climate change projections to be useful, the magnitude of change must be understood relative to the magnitude of uncertainty in model predictions. We quantified the signal-to-noise ratio in projected distributional responses of boreal birds to climate change, and compared sources of uncertainty. Boosted regression tree models of abundance were generated for 80 boreal-breeding bird species using a comprehensive data set of standardized avian point counts (349 629 surveys at 122 202 unique locations) and 4-km climate, land use, and topographic data. For projected changes in abundance, we calculated signal-to-noise ratios and examined variance components related to choice of global climatemodel (GCM) and two sources of species distribution model (SDM) uncertainty: sampling error and variable selection. We also evaluated spatial, temporal, and interspecific variation in these sources of uncertainty. The mean signal-to-noise ratio across species increased over time to 2.87 by the end of the 21st century, with the signal greater than the noise for 88% of species. Across species, climate change represented the largest component (0.44) of variance in projected abundance change. Among sources of uncertainty evaluated, choice of GCM (mean variance component = 0.17) was most important for 66% of species, sampling error (mean = 0.12) for 29% of species, and variable selection (mean = 0.05) for 5% of species. Increasing the number of GCMs from four to 19 had minor effects on these results. The range of projected changes and uncertainty characteristics across species differed markedly, reinforcing the individuality of species' responses to climate change and the challenges of one-size-fits-all approaches to climate change adaptation. We discuss the usefulness of different conservation approaches depending on the strength of the climate change signal relative to the noise, as well as the dominant source of prediction uncertainty.
Article
Lightning plays an important role in atmospheric chemistry and in the initiation of wildfires, but the impact of global warming on lightning rates is poorly constrained. Here we propose that the lightning flash rate is proportional to the convective available potential energy (CAPE) times the precipitation rate. Using observations, the product of CAPE and precipitation explains 77% of the variance in the time series of total cloud-to-ground lightning flashes over the contiguous United States (CONUS). Storms convert CAPE times precipitated water mass to discharged lightning energy with an efficiency of 1%. When this proxy is applied to 11 climate models, CONUS lightning strikes are predicted to increase 12 ± 5% per degree Celsius of global warming and about 50% over this century.
Article
[1] In this study, components of the Max Planck Institute Earth System Model (MPI-ESM) were used to explore how changes in lightning induced by climate change alter wildfire activity. To investigate how climate change alters global flash frequency simulations with the atmospheric general circulation model ECHAM6 were performed for the time periods pre-industrial, present-day and three future scenarios. The effect of changes in lightning activity on fire occurence was derived from simulations with the land surface vegetation model JSBACH. Global cloud-to-ground (CG) lightning activity decreased by 3.3% under pre-industrial climate and increased by up to 21.3% for the RCP85 projection at the end of the century when compared to present-day, respectively. Relative changes were most pronounced in North America and Northeastern Asia. Global burned area was little affected by these changes and only increased by up to 3.3% for RCP85. However, on the regional scale, significant changes occured. For instance, burned area increases of over 100% were found in high latitude regions while also several regions were identified where burned area declined, such as parts of South America and Africa.
Article
Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. The generalization error for forests converges a.s. to a limit as the number of trees in the forest becomes large. The generalization error of a forest of tree classifiers depends on the strength of the individual trees in the forest and the correlation between them. Using a random selection of features to split each node yields error rates that compare favorably to Adaboost (Y. Freund & R. Schapire, Machine Learning: Proceedings of the Thirteenth International conference, ***, 148–156), but are more robust with respect to noise. Internal estimates monitor error, strength, and correlation and these are used to show the response to increasing the number of features used in the splitting. Internal estimates are also used to measure variable importance. These ideas are also applicable to regression.
Article
In this paper we propose two ways to deal with the imbalanced data classification problem using random forest. One is based on cost sensitive learning, and the other is based on a sampling technique. Performance metrics such as precision and recall, false positive rate and false negative rate, F-measure and weighted accuracy are computed. Both methods are shown to improve the prediction accuracy of the minority class, and have favorable performance compared to the existing algorithms.
Article
An integrated forecasting and data assimilation system has been and is continuing to be developed by the Meteorological Research Branch (MRB) in partnership with the Canadian Meteorological Centre (CMC) of Environment Canada. Pan I of this two-part paper motivates the development of the new system, summarizes various considerations taken into its design, and describes its main characteristics.
Article
Of 16 documented lightning discharges 11 caused forest fires in western Montana forests. All 11 discharges exhibited a long-continuing current (LCC) phase of at least 40 msec duration. Of the 5 non-fire discharges, 2 had LCC phases and 3 did not. Results to date strongly support the hypothesis that forest fires are caused by discharges with long-continuing currents. However, available data suggest that discrete discharges (cloud-to-ground discharges without LCC portions) cannot be entirely ruled out as a source of forest fuel ignition.
Article
The party package (Hothorn, Hornik, and Zeileis 2006) aims at providing a recursive part(y)itioning laboratory assembling various high- and low-level tools for building tree-based regression and classification models. This includes conditional inference trees (ctree), condi- tional inference forests (cforest) and parametric model trees (mob). At the core of the pack- age is ctree, an implementation of conditional inference trees which embed tree-structured regression models into a well defined theory of conditional inference procedures. This non- parametric class of regression trees is applicable to all kinds of regression problems, including nominal, ordinal, numeric, censored as well as multivariate response variables and arbitrary measurement scales of the covariates. This vignette comprises a practical guide to exploiting the flexible and extensible computational tools in party for fitting and visualizing conditional inference trees.
Article
The hypothesis that lightning fires are caused by a special type of lightning discharge has been presented several times in literature over the past 20 years. Working with laboratory sparks, McEachron and Hagenguth [1942] suggested that ignition by natural lightning is usually caused by a discharge having a long-continuing current phase. This view is widely accepted [Berger, 1947; Malan, 1963; Loeb, 1966], even though field corroboration has been almost nil. To the authors' knowledge, the two discharges and resultant fires documented by Norinder et al. [1958] are the only natural events reported to date wherein both the discharge and its ignition effects were documented. Both of those events exhibited long-continuing current phases. Employing the instruments and techniques described below, we have recently documented characteristics of seven lightning discharges that caused forest fires in western Montana. All seven discharges had a longcontinuing current phase. To document the effects of specific discharges on forest fuels, we observed the ground terminals of the discharges (usually trees) from a light aircraft flying 2 to 4 km in a'dvance of the storm. Immediately on seeing a nearby discharge, the aerial observer transmitted a discharge identification number to the lightning recording stations over air-to-ground radio. His voice transmission was recorded on the same magnetic tape that recorded the field change cause'd by the discharge. The field change was thus linked to the aerial observation of the discharge. The aircraft then circled in the immediate vicinity of the ground terminal, and the observer attempted to locate the ground terminal visually. If he was successful, he photographed the terminal and noted its location for subsequent on-the-ground study of lightning effects. Using the sampling procedures described above, the aerial observer attempted to identify the ground terminals of thirty observed discharges. He was able to identify seven of them. All seven groun'd terminals exhibited fire, and all the fires were caused by discharges having at least one long-continuing current phase exceeding 40-msec duration. Of the 856 cloud-toground discharges recorded during 1965 and 1966, about half contained a long-continuing current phase. The lightning sensors located at the recording stations are fast and slow field change meters and a nondirectional luminosity device. The output of the sensors is recorded on a 7-channel tape recorder. The data are retrieved for analysis on a 14-channel oscillograph. The fast and slow field change sensors are identical, with the exception of the grid resistor in the electrometer stage. The grid resistor establishes the sensor time constant, which is 5 msec for the fast antenna and 5 sec for the slow antenna. The field change sensor circuit consists essentially of a fiat-plate antenna, an electrometer stage, an emitter follower, and an amplifier. The output is directly proportional to the field change at the antenna. The luminosity measuring system consists of a cone-shaped mirror and a parabolic mirror with an uncooled lead sulfide detector at the focal point.. The detector is capacitor-coupled
Article
Lightning ground flash and stroke observations were made with a single‐station gated, wideband magnetic direction‐finding system with a nominal range of 180 km located in Southern Ontario during the May‐September lightning‐active seasons of 1982 and 1983. Information was recorded on the azimuth of arrival, time, amplitude, stroke multiplicity and order, and polarity. The local climatology and seasonal statistics of lightning are analysed and summarized, and compared with standard observations of thunderstorm days and hours. Regional daily flashing rates and extremes for periods of 5 to 60 min were found to have a good empirical relationship. About 15% of the flashes had multiple strokes, generally less than 10 but with as many as 14 strokes. About 8% of the flashes were positive discharges; 3% of these were multistroke with no more than 2 strokes.The lightning activity exhibits well defined diurnal peaks in the afternoon and at night 1–3 h before sunrise. The time interval between strokes was found to have a lognormal distribution with modal and median values of 60 and 75 ms, respectively, and no significant dependence on the order of stroke. The stroke‐to‐stroke amplitude changes within the same flash show that subsequent stroke amplitudes are often greater than the first. Subsequent strokes follow many patterns of change, the most common being an amplitude oscillating with ascending stroke order. The multistroke flash duration median values rose from about 80 ms for 2‐stroke flashes to about 650 ms for 8‐stroke flashes.Under certain assumptions of system detection efficiency and range limits a regional ground flash density of 1.62 and 2.44 km a was estimated for the two years.
Article
The coupling of synoptic scale weather conditions with local scale weather and fuel conditions was examined for 2551 fires and 1 537 624 lightning strikes for the May through August fire seasons in 1988, 1989, 1992, and 1993 in Alberta and Saskatchewan. The probability of lightning fire occurrence (number of fires/number of strikes) is near zero until the Fine Fuel Moisture Code reaches 87 (moisture content of 14% dry weight), after which the probability increases rapidly. Duff Moisture and Drought Codes show less clear increases. In all cases, the probability of fire occurrence was low (the number of strikes greatly exceeded the number of forest fires), suggesting that lightning fire ignition coupled with early spread to detection was an uncommon event. This low probability of fire occurrence even at low fuel moisture may be a result of the arrangement and continuity of fuels in the boreal and subalpine forests. The literature suggests a higher probability of lightning-ignited fires in qualitatively different fuels, e.g., grasslands. The higher probability of fire at lower fuel moistures occurred primarily when high pressure dominated (positive 50-kPa anomaly) for at least 3 days and less than 1.5 mm precipitation occurred. The highest number of lightning strikes and largest number of fires also occurred when high pressure dominated. The high lightning numbers during high pressure systems were logistically related to increasing atmospheric instability (K-index).
Article
Boreal forest dynamics are largely driven by disturbance, and fire is a prevalent force of change across the boreal circumpolar region. North American and Eurasian boreal fire regimes are known to be very different but there are few quantitative comparison studies. Russian and Canadian boreal fire regimes are compared using fire weather, fire statistics, fire behaviour, and C emissions data from two large study areas. Fuel consumption, head fire intensity, and C emissions were modelled using fire weather data, fuels data and burned area polygons for all large (200+ ha) fires that occurred in the study areas during 2001–2007. Fire behaviour and C emissions of each large fire were simulated with the Canadian Fire Effects Model (CanFIRE) using fuel type and fuel load data of the burned areas, and Canadian Forest Fire Weather Index System parameters, as interpolated to the fire from the weather station network on the average active fire date. In the Russian study area located in central Siberia, there was an annual average of 1441.9 large fires per 100 M ha of forest land that burned 1.89 M ha (average large fire size = 1312 ha, mean fire return interval = 52.9 years) with an average fire intensity of 4858 kW m−1. In the western Canada study area, there was an annual average of 93.7 large fires per 100 M ha of forest land that burned 0.56 M ha of forest (average large fire size = 5930 ha, mean fire return interval = 179.9 years) with an average fire intensity of 6047 kW m−1. The 2001–2007 fire size distribution and annual area burned in the Canadian study area were very similar to 1970–2009 statistics, although large fire frequency was higher and average large fire size was smaller. Similar long-term fire statistics for Russia currently do not exist for comparison. The C emissions rate (t ha−1 of burned area) was 53% higher in the Canadian study area due to higher pre-burn forest floor fuel loads and higher fuel consumption by crown fires. However, the Russian study area had much higher total C emissions (per 100 M ha of forest area) because of greater annual area burned. The Russian C emissions estimate in this study is likely conservative due to low forest floor fuel load estimates in available datasets. Fire regime differences are discussed in terms of fuel, weather, and fire ecology.
Article
A simplified procedure is described for computation of equivalent potential temperature which remains valid in situations such as in the tropics where a term which is omitted in the derivation of the conventional formula can lead to an error of several degrees absolute. -from Author
Article
Composite profiles of temperature, humidity and wind are constructed for severe thunderstorms that formed over central Alberta, Canada, during the period 1967–2000. Storms were divided into three categories consisting of 13 non-tornadic storms which produced hail ≥ 3 cm in diameter but no reported tornadoes, 61 weak tornadoes (F0 and F1) and 13 significant tornadoes (F2–F4). All three composites showed potential instability through most of the sounding. However, thermodynamic parameters did not discriminate among the three categories. In contrast, however, composite hodographs did show noticeable differences among the three groups. The hodograph for the significant tornado composite exhibited strong low-level veering winds having 0–3 km helicity of 65 m s and southwest winds of about 20 m s in the mid- to upper levels. The hodograph for the weak tornado composite showed weak low-level veering winds (0–3 km helicity of 17 m s) and much lighter winds overall compared with the other two composites. The non-tornado composite hodograph had nearly unidirectional southwest winds of about 15 m s in the mid- to upper levels. The significant tornadic storms were similar to typical mid-latitude supercell hodographs.
Article
This study presents a method for estimating daily rainfall on a 0.05° latitude/longitude grid covering all of New Zealand for the period 1960–2004 using a second order derivative trivariate thin plate smoothing spline spatial interpolation model. Use of a hand-drawn (and subsequently digitised) mean annual rainfall surface as an independent variable in the interpolation is shown to reduce the interpolation error compared with using an elevation surface. This result is confirmed when long-term average annual rainfall data, derived from the daily interpolations, are validated using long-term river flow data. Copyright © 2006 Royal Meteorological Society
Article
1. Bioclimate envelope models are widely used to predict the potential distribution of species under climate change, but they are conceptually also suitable to match policies and practices to anticipated or observed climate change, for example through species choice in reforestation. Projections of bioclimate envelope models, however, come with large uncertainties due to different climate change scenarios, modelling methods and other factors. 2 In this paper we present a novel approach to evaluate uncertainty in model-based recommendations for natural resource management. Rather than evaluating variability in modelling results as a whole, we extract a particular statistic of interest from multiple model runs, e.g. species suitability for a particular reforestation site. Then, this statistic is subjected to analysis of variance, aiming to narrow the range of projections that practitioners need to consider. 3. In four case studies for western Canada we evaluate five sources of uncertainty with two to five treatment levels, including modelling methods, interpolation type for climate data, inclusion of topo-edaphic variables, choice of general circulation models, and choice of emission scenarios. As dependent variables, we evaluate changes to tree species habitat and ecosystem distributions under 144 treatment combinations. 4. For these case studies, we find that the inclusion of topo-edaphic variables as predictors reduces projected habitat shifts by a quarter, and general circulation models had major main effects. Our contrasting modelling approaches primarily contributed to uncertainty through interaction terms with climate change predictions, i.e. the methods behaved differently for particular climate change scenarios (e.g. warm & moist scenarios) but similar for others. 5. Synthesis and applications. Partitioning of variance components helps with the interpretation of modelling results and reveals how models can most efficiently be improved. Quantifying variance components for main effects and interactions among sources of uncertainty also offers researchers the opportunity to filter out biologically and statistically unreasonable modelling results, providing practitioners with an improved range of predictions for climate-informed natural resource management.
Article
Lightning in all corners of the world is monitored by one or more land- or space-based lightning locating systems (LLSs). The applications that have driven these developments are numerous and varied. This paper describes the history leading to modern LLSs that sense lightning radiation fields at multiple remote sensors, focusing on the interactions between enabling technology, scientific discovery, technical development, and uses of the data. An overview of all widely used detection and location methods is provided, including a general discussion of their relative strengths and weaknesses for various applications. The U.S. National Lightning Detection Network (NLDN) is presented as a case study, since this LLS has been providing real-time lightning information since the early 1980s, and has provided continental-scale (U.S.) information to research and operational users since 1989. This network has also undergone a series of improvements during its >20-year life in response to evolving detection technologies and expanding requirements for applications. Recent analyses of modeled and actual performance of the current NLDN are also summarized. The paper concludes with a view of the short- and long-term requirements for improved lightning measurements that are needed to address some open scientific questions and fill the needs of emerging applications.
Conference Paper
Existing volume visualization techniques are typically applied to a three-dimensional grid. This presents some challenging problems in the visualization of environmental data. This data often consists of unevenly distributed samples. Typically a two-step approach is used to visualize environmental data. First the unevenly distributed sample data are modeled onto a uniform 3-D grid. This grid model is subsequently rendered using conventional grid-based visualization techniques. This paper discusses some of the limitations of this approach and highlights areas where further research is needed to improve the accuracy of visualization for environmental applications.
Article
Classification and feature selection of genomics or transcriptomics data is often hampered by the large number of features as compared with the small number of samples available. Moreover, features represented by probes that either have similar molecular functions (gene expression analysis) or genomic locations (DNA copy number analysis) are highly correlated. Classical model selection methods such as penalized logistic regression or random forest become unstable in the presence of high feature correlations. Sophisticated penalties such as group Lasso or fused Lasso can force the models to assign similar weights to correlated features and thus improve model stability and interpretability. In this article, we show that the measures of feature relevance corresponding to the above-mentioned methods are biased such that the weights of the features belonging to groups of correlated features decrease as the sizes of the groups increase, which leads to incorrect model interpretation and misleading feature ranking. With simulation experiments, we demonstrate that Lasso logistic regression, fused support vector machine, group Lasso and random forest models suffer from correlation bias. Using simulations, we show that two related methods for group selection based on feature clustering can be used for correcting the correlation bias. These techniques also improve the stability and the accuracy of the baseline models. We apply all methods investigated to a breast cancer and a bladder cancer arrayCGH dataset and in order to identify copy number aberrations predictive of tumor phenotype. R code can be found at: http://www.mpi-inf.mpg.de/~laura/Clustering.r.
Article
Thesis (M. Sc.)--University of Alberta, 1991. Includes bibliographical references (p. 79-82). Photocopy.