No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Assessing Post-Fire Regeneration in a Mediterranean Mixed Forest Using Lidar Data and Artifi cial Neural Networks
Abstract
Post-wildfi re management practices can greatly infl uence vegetation
condition and dynamics, and are crucial in Mediterranean
erosion-prone poor soil sites. Acquiring accurate ground
inventory data is time-consuming, expensive and limited to
small areas; but lidar data can be used to assess the impact
of fi res, and also to determine the convenient silvicultural
measurements which should be carried out for site restoration.
The aim of this paper was to assess the post-fi re regeneration
status of the vegetation in Sant Llorenç del Munt massif
after a wildfi re in summer 2003 by modeling the relationship
between lidar height bins and canopy height model (CHM) with
fi eld data. Artifi cial Neural Network (ANN) prediction models
provided estimations of vegetation fraction cover, average
height (HM) over 1.30 m and number of stems over 1.30 m,
with Pearson r values between 0.18 and 0.83. Classifi cation
models built with the same variables allowed separating two
ground-based regeneration classes (good and scarce regeneration)
with an approximate accuracy of 83 to 76 percent
(model building and validation data, respectively).
... The anticipated complexity of modeling use of firefighting resources, and the fact that some successful applications had been developed before for other fire problems, led us to select artificial neural networks (ANNs) as a modeling technique. ANNs have been successfully applied to problems such as fire occurrence prediction (Vega-Garcia et al. 1996, Vasconcelos et al. 2001, Li et al. 2009, Vasilakos et al. 2009, Karouni et al. 2014, regional forest fire susceptibility (Dimuccio et al. 2011), forest fire risk prediction and firefighting management in Galicia (Alonso-Betanzos et al. 2003), burned area mapping (Mitrakis et al. 2012), fire-landscape structure relations (Vega-Garcia & Chuvieco 2006, Ruiz-Mirazo et al. 2012, and the evaluation of forest regeneration after fire (Debouk et al. 2013). ANN models are a reliable alternative to traditional statistical methods because they are robust pattern detectors even for unpredictable non-linear relationships (Scrinzi et al. 2007), they are not affected by multicollinearity or non-normal distributions (Hilbert & Ostendorf 2001) like statistical techniques, and they are flexible in terms of structure. ...
... software (NeuralWare 2009). The models were computed as in Alcázar et al. (2008) and Debouk et al. (2013), but with three output nodes (one for each of the suppression dependent variables). ...
In Spain, the established fire control policy states that all fires must be controlled and put out as soon as possible. Though budgets have not restricted operations until recently, we still experience large fires and we often face multiple-fire situations. Furthermore, fire conditions are expected to worsen in the future and budgets are expected to drop. To optimize the deployment of firefighting resources, we must gain insights into the factors affecting how it is conducted. We analyzed the national data base of historical fire records in Spain for patterns of deployment of fire suppression resources for large fires. We used artificial neural networks to model the relationships between the daily fire load, fire duration, fire type, fire size and response time, and the personnel and terrestrial and aerial units deployed for each fire in the period 1998-2008. Most of the models highlighted the positive correlation of burned area and fire duration with the number of resources assigned to each fire and some highlighted the negative influence of daily fire load. We found evidence suggesting that firefighting resources in Spain may already be under duress in their compliance with Spain's current full suppression policy.
... Only few studies focused on forest regeneration characterisation using ALS data: understory coverage estimation (Latifi et al. 2017, Venier et al. 2019, post-fire vegetation characterization (Martin-Alcon et al. 2015), regeneration stems density and height estimation (Debouk et al. 2013, Imangholiloo et al. 2020. ...
... In addition, studies have shown that fire driving factors have a nonlinear relationship with forest fires, so nonparametric models are more suitable than parametric models for forest fire prediction and variable analysis [30]. To date, many scholars have used artificial neural networks (ANNs) to classify fire hazard levels [31] and to assess ecological restoration after a fire [32,33], and some scholars have also used this method to study the area of fire spread. However, the combination of multiple forest fire factors and the use of ANNs to predict forest fires in China has not yet been supported by mature research. ...
Although low-intensity forest fires are a necessary part of healthy echo system, high-intensity forest fires continue to affect the diversity of forest ecosystems and species. Therefore, it is necessary to study the driving factors of forest fires and explore their possible locations and probabilities in complex forest terrain. In this article, we determined the relative influences of different types of factors on the occurrence of forest fires in Heilongjiang forest areas and compared the performance of artificial neural networks and logistic regression for wildfire prediction. By analyzing Heilongjiang forest fire data from 2002 to 2015 and constructing a model, we found that climate factors, topographical factors, and vegetation type factors play a crucial role in Heilongjiang’s wildfires. During the fire prevention period, temperature and wind speed have a more significant influence than other factors. According to the variable screening that we conducted, the model built by the variables that we used can predict 84% of forest fires in Heilongjiang Province. For recent wildfires (2019–2020) in most areas, we can use artificial neural networks for relatively accurate verification (85.2%). Therefore, artificial neural networks are very suitable for the prediction of forest fires in Heilongjiang Province. Through the prediction results, we also created a probability distribution map of fire occurrence in the study area. On this basis, we also analyzed the changes in the probability of natural fires under the weather changing trend, which can effectively aid in fire prevention and extinguishment.
... Airborne laser scanning (ALS) has grown in popularity for its ability to provide landscape-scale data related to forest structure, including vegetation height and stem volume (Naesset et al. 2004;Vastaranta et al. 2011). ALS may be employed in monitoring forest condition following harvest (Wulder et al. 2008) and wildfire (Bolton et al. 2015), both achievable objectives even with the low-density ALS data commonly collected in the 2000s (Naesset et al. 2004;Debouk et al. 2013). The availability of ALS data provides an excellent opportunity to assess the state of the aging network of conventional seismic lines in Alberta on a large scale, without establishing expensive and often logistically difficult field campaigns. ...
Linear disturbances from geological exploration (i.e., seismic lines) have an extensive footprint across much of Canada’s western boreal forest; however, how seismic lines interact with subsequent wildfire remains poorly understood. We assessed whether wildfires effectively mitigate the footprint of seismic lines by promoting forest recovery. We evaluated the forest structure of legacy seismic lines burned in 2001 and 2002 by comparing them against adjacent unburned control plots, using metrics derived from airborne laser scanning (ALS) data collected between 2007 and 2009. ALS metrics identified persistent differences in vegetation height and structure between seismic lines and adjacent forest controls, 7–9 years post-fire. Median canopy height was 2.90 m lower on upland seismic lines and 1.94 m lower on lowland seismic lines than on adjacent controls, corresponding to a 21% and 25% height reduction, respectively. Field surveys revealed greater graminoid and nonvascular plant cover, and lower tall-tree cover and dead vegetative matter, on seismic lines, in comparison with controls. Our results show that tree recovery remains significantly delayed on most upland and lowland burned seismic lines in our study area, and that additional management efforts or longer timescales may be required to restore these fragmented landscapes.
... Previous studies have employed ANNs to investigate fire danger (Bisquert et al., 2012;Pai et al., 2020), wildfire vulnerability (Dimuccio et al., 2011;de Bem et al., 2019), wildfire risk assessment (Li et al., 2009;Jafari Goldarag et al., 2016;Lall and Mathibela, 2016), burned area detection (Maeda et al., 2009;Gómez and Martín, 2011), pre-and post-fire vegetation (Debouk et al., 2013;Polinova et al., 2019), causes of wildfires (Rodrigues and de la Riva, 2014), and flame and smoke detection (Chetehouna et al., 2015;Hossain et al., 2019). However, the literature currently does not provide a satisfactorily large number of studies in which ANNs have been employed to estimate the probability of wildfire occurrence. ...
Wildfires are a major disturbance in the Mediterranean Basin and an ecological factor that constantly alters the landscape. In this context, it is crucial to understand where wildfires are more likely to occur as well as the drivers guiding them in complex landscapes such as the Mediterranean area. The objectives of this study are to estimate wildfire probability occurrence as a function of biophysical and human-related drivers, to provide an assessment of the relative impact of each driver and analyze the performance of machine learning techniques compared to traditional regression modeling. By employing an Artificial Neural Network model and fire data (2004–2012), we estimated wildfire probability across two geographical regions covering most of the Italian territory: Alpine and subalpine region and Insular and peninsular region. The high classification accuracy (0.68 for the Alpine and subalpine region and 0.76 for the Insular and peninsular region) and good performances of the technique (AUC values of 0.82 and 0.76, respectively) suggest that our model can be used in the areas studied to assess wildfire probability occurrence. We compared our model with a logistic function, which showed a weaker
predictive power (AUC values of 0.78 for the Alpine and subalpine region and 0.65 for the Insular and peninsular region) compared to the Artificial Neural Network. In addition, we assessed the importance of each variable by isolating it in the model. The importance of an individual variable differed between the two regions, underscoring the high diversity of wildfire occurrence drivers in Mediterranean landscapes. Results show that in the
Alpine and subalpine region, the presence of forest is the most important variable, while climate resulted as being the most important variable in the Insular and peninsular region. The majority of areas recently affected by large wildfires in both regions have been correctly classified by the ANN model as ‘high fire probability’. Hence, the use of an Artificial Neural Network is efficient and robust for understanding the probability of wildfire occurrence in Italy and other similar complex landscapes.
... ALS data from the Spanish National Plan for Aerial Orthophotography (PNOA) have yielded good results for monitoring post-fire recovery and forest structural diversity, either using regressions (Kane et al. 2013), or classifications using ALS data alone (Debouk, Riera-Tatché, and Vega-Garcia 2013) or in combination with optical data (Martín-Alcón et al. 2015). New approaches, such as the one developed by Listopad et al. (2015), have emerged to provide estimations of vegetation structural diversity based on the proportion of returns in different canopy height ranges. ...
Forest fires can change forest structure and composition, and low-density Airborne Laser Scanning (ALS) can be a valuable tool for evaluating post-fire vegetation response. The aim of this study is to analyze the structural diversity differences in Mediterranean Pinus halepensis Mill. forests affected by wildfires on different dates from 1986 to 2009. Several types of ALS metrics, such as the Light Detection and Ranging (LiDAR) Height Diversity Index (LHDI), the LiDAR Height Evenness Index (LHEI), and vertical and horizontal continuity of vegetation, as well as topographic metrics, were obtained in raster format from low point density data. In order to map burned and unburned areas, differentiate fire occurrence dates, and distinguish between old and more recent fires, a sample of pixels was previously selected to assess the existence of differences in forest structure using the Kruskal-Wallis test. Then, k-nearest neighbors algorithm (k-NN), support vector machine (SVM) and random forest (RF) classifiers were compared to select the most accurate technique. The results showed that, in more recent fires, around 70% of the laser returns came from grass and shrub layers, yielding low LHDI and LHEI values (0.37-0.65 and 0.28-0.46, respectively). In contrast, the areas burned more than 20 years ago had higher LHDI and LHEI values due to the growth of the shrub and tree strata. The classification of burned and unburned areas yielded an overall accuracy of 89.64% using the RF method. SVM was the best classifier for identifying the structural differences between fires occurring on different dates, with an overall accuracy of 68.79%. Furthermore, SVM yielded an overall accuracy of 75.49% for the classification between old and more recent fires. ARTICLE HISTORY
... software [47]. The models were computed as in Alcázar et al. [48] and Debouk et al. [49]. The cascade-correlation algorithm starts model building with the simplest possible architecture, adding nodes until an optimal multi-layer structure is reached [46]. ...
Wildfires are currently considered the major threat to forests in Mediterranean countries. It has been implied that a large percentage of arson-caused fires in Spain are connected with the extensive reforestation programs implemented between 1940 and 1970. However, no consistent studies have been conducted to study the relationships between arson-caused fires and stand origin. Therefore, the goal of this study was to analyze occurrences and model the influence of forest stand origin (artificial or not) on the development of wildfires in peninsular Spain. Twenty-one neural network models were trained to estimate fire incidence through fire type (surface or crown fire), burned area and total treed burned area, based on stand age (years), canopy cover (%), natural age class (from seedling to mature stages) and fuel type classification. Models were built for reforested stands and natural stands of Pinus pinaster Ait., the Mediterranean pines Pinus sylvestris L., Pinus nigra Arn., Pinus halepensis Mill. and Eucalyptus sp. L'Hér., or groups of these species, and the resulting models were compared. Reforested stands presented higher fire incidence than natural stands mainly for productive species like Pinus pinaster Ait. According to the fire type models, thickets had a large influence in the development of crown fires in reforested stands in a general model for all species, the model with the Mediterranean group of pines, and the Pinus pinaster Ait. model. Vertical continuity influenced crown fire propagation in natural Mediterranean pines and in Eucalyptus stands. Presence of shrubs, grasslands and wood slash was related to surface fires in models for both reforested and natural stands. The results suggested that stand origin was influential on fire incidence, at least with regard to fire type and commercial species in the northwestern region of Spain.
... For example, Marino et al. (2017b) compared metrics derived from < 4 m strata returns of three LiDAR datasets (1 pulse × m -2 ) acquired pre-(2011) and post-fire (2012,2014) in Garajonay and characteri zed vegetation recovery, demonstrating the value of repetitive LiDAR acquisitions. Debouk et al. (2013) employed low density LiDAR data (0.7 pulse × m -2 ) ac quired over 104 km 2 of mixed forest (P. halepensis and Q. ilex) in Barcelona five years after fire, and modelled vegetation recovery with an Artificial Neural Network (ANN) for classification and mapping. ...
Forest ecosystems serve a host of services and societal benefits, including carbon storage, habitat for fauna, recreation, and provision of wood or non-wood products. In a context of complex demands on forest resources, identifying priorities for biodiversity and carbon budgets require accurate tools with sufficient temporal frequency. Moreover, understanding long term forest dynamics is necessary for sustainable planning and management. Remote sensing (RS) is a powerful means for analysis, synthesis and report, providing insights and contributing to inform decisions upon forest ecosystems. In this communication we review current applications of RS techniques in Spanish forests, examining possible trends, needs, and opportunities offered by RS in a forestry context. Currently, wall-to-wall optical and LiDAR data are extensively used for a wide range of applications—many times in combination—whilst radar or hyperspectral data are rarely used in the analysis of Spanish forests. Unmanned Aerial Vehicles (UAVs) carrying visible and infrared sensors are gaining ground in acquisition of data locally and at small scale, particularly for health assessments. Forest fire identification and characterization are prevalent applications at the landscape scale, whereas structural assessments are the most widespread analyses carried out at limited extents. Unparalleled opportunities are offered by the availability of diverse RS data like those provided by the European Copernicus programme and recent satellite LiDAR launches, processing capacity, and synergies with other ancillary sources to produce information of our forests. Overall, we live in times of unprecedented opportunities for monitoring forest ecosystems with a growing support from RS technologies.
... In some shrub species, there is a strong link between shrub height and other biophysical characteristics (e.g., cover, AGB, canopy volume[27]), thus making Lidar advantageous for vegetation structure measurements. Metrics derived from Lidar (e.g., mean height, variance of height, canopy relief ratio) can be correlated with biophysical vegetation characteristics in the field using statistical methods such as Classical Multiple Linear Regression (CMLR)[28], Partial Least Square Regression[29], Hierarchical Bayesian[30], Random Forests[31], and Artificial Neural Networks[32]. The machine learning algorithm Random Forests (RF) assembles the analysis of Classification and Regression Trees (CART) by bootstrapping samples to iteratively construct a large number of decision trees, each grown with a randomized subset of predictors[33]. ...
Our study objectives were to model the aboveground biomass in a xeric shrub-steppe landscape with airborne light detection and ranging (Lidar) and explore the uncertainty associated with the models we created. We incorporated vegetation vertical structure information obtained from Lidar with ground-measured biomass data, allowing us to scale shrub biomass from small field sites (1 m subplots and 1 ha plots) to a larger landscape. A series of airborne Lidar-derived vegetation metrics were trained and linked with the field-measured biomass in Random Forests (RF) regression models. A Stepwise Multiple Regression (SMR) model was also explored as a comparison. Our results demonstrated that the important predictors from Lidar-derived metrics had a strong correlation with field-measured biomass in the RF regression models with a pseudo R2 of 0.76 and RMSE of 125 g/m2 for shrub biomass and a pseudo R2 of 0.74 and RMSE of 141 g/m2 for total biomass, and a weak correlation with field-measured herbaceous biomass. The SMR results were similar but slightly better than RF, explaining 77-79% of the variance, with RMSE ranging from 120 to 129 g/m2 for shrub and total biomass, respectively. We further explored the computational efficiency and relative accuracies of using point cloud and raster Lidar metrics at different resolutions (1 m to 1 ha). Metrics derived from the Lidar point cloud processing led to improved biomass estimates at nearly all resolutions in comparison to raster-derived Lidar metrics. Only at 1 m were the results from the point cloud and raster products nearly equivalent. The best Lidar prediction models of biomass at the plot-level (1 ha) were achieved when Lidar metrics were derived from an average of fine resolution (1 m) metrics to minimize boundary effects and to smooth variability. Overall, both RF and SMR methods explained more than 74% of the variance in biomass, with the most important Lidar variables being associated with vegetation structure and statistical measures of this structure (e.g., standard deviation of height was a strong predictor of biomass). Using our model results, we developed spatially-explicit Lidar estimates of total and shrub biomass across our study site in the Great Basin, U.S.A., for monitoring and planning in this imperiled ecosystem.
... In contrast to passive sensors measuring the sun's reflected radiation these sensors supply their own power. LiDAR data have demonstrated potential for characterising fire-induced changes in overstorey vegetation characteristics (Wang and Glenn 2009;Wulder et al. 2009;Kwak et al. 2010;Magnussen and Wulder 2012), For example, LiDAR has been used for post-fire assessments to quantify tree regeneration (Debouk et al. 2013), assess how post-fire forest structure varies with burn severity (Kane et al. 2013), and estimate post-fire tree height (Magnussen and Wulder 2012). However, operational use of LiDAR remains limited given the sparse coverage of pre-fire data and the expense of acquiring new post-fire data. ...
Comprehensive assessment of ecological change after fires have burned forests and rangelands is important if we are to understand, predict and measure fire effects. We highlight the challenges in effective assessment of fire and burn severity in the field and using both remote sensing and simulation models. We draw on diverse recent research for guidance on assessing fire effects on vegetation and soil using field methods, remote sensing and models. We suggest that instead of collapsing many diverse, complex and interacting fire effects into a single severity index, the effects of fire should be directly measured and then integrated into severity index keys specifically designed for objective severity assessment. Using soil burn severity measures as examples, we highlight best practices for selecting imagery, designing an index, determining timing and deciding what to measure, emphasising continuous variables measureable in the field and from remote sensing. We also urge the development of a severity field assessment database and research to further our understanding of causal mechanisms linking fire and burn severity to conditions before and during fires to support improved models linking fire behaviour and severity and for forecasting effects of future fires.
... In addition, the provision of continuous spatially-explicit information at the 426 landscape scale allows easy implementation of relevant forestry applications such as the 427 characterization and mapping of forest canopy fuels and fire risk (Erdody and Moskal 2010;428 García et al. 2011;González-Olabarria et al. 2012;Mutlu et al. 2008;Pierce et al. 2012;Riaño 429 et al. 2007). The resulting maps of regeneration types also offer a starting point for the analysis 430 of post-disturbance vegetation dynamics at large scales (Debouk et al. 2013;Goetz et al. 2010;431 Holmgren et al. 2008). Our methodology is appropriate for large-scale assessment, since 432 regional and national LiDAR coverages are often collected at low-resolution. ...
Wildfires play a major role in driving vegetation changes and can cause important environmental and economic losses in Mediterranean forests, especially where the dominant species lacks efficient postfire regeneration mechanisms. In these areas, postdisturbance vegetation management strategies need to be based on detailed, spatially continuous inventories of the burned area. Here, we present a methodology in which we combine airborne LiDAR and multispectral imagery to assess postfire regeneration types in a spatially continuous way, using a Mediterranean black pine (Pinus nigra Arn ssp. salzmannii) forest that burned in 1998 as a case study. Five postfire regeneration types were obtained by clustering field-plot data using Ward’s method. Two of the five regeneration types presented high tree cover (one clearly dominated by hardwoods and the other dominated by pines), a third type presented low to moderate tree cover, being dominated by hardwoods, and the remaining two types matched to areas dominated by soil–herbaceous or shrub layers with very low or no tree cover (i.e., very low to no tree species regeneration). These five types of regeneration were used to conduct a supervised classification of remote sensing data using a nonparametric supervised classification technique. Compared with independent field validation points, the remote sensing based assessment method resulted in a global classification accuracy of 82.7%. Proportions of regeneration types in the study area indicated a general shift from the former pine-dominated forest toward hardwood dominance and showed no serious problems of regeneration failure. Our methodological approach appears to be appropriate for informing postdisturbance vegetation management strategies over large areas.
Wildfires are a critical driver of landscape transformation on Earth, representing a dynamic and ephemeral process that poses challenges for accurate early detection. To address this challenge, researchers have increasingly turned to deep learning techniques, which have demonstrated remarkable potential in enhancing the performance of wildfire detection systems. This paper provides a comprehensive review of fire detection using deep learning, spanning from 1990 to 2023. This study employed a comprehensive approach, combining bibliometric analysis, qualitative and quantitative methods, and systematic review techniques to examine the advancements in fire detection using deep learning in remote sensing. It unveils key trends in publication patterns, author collaborations, and thematic focuses, emphasizing the remarkable growth in fire detection using deep learning in remote sensing (FDDL) research, especially from the 2010s onward, fueled by advancements in computational power and remote sensing technologies. The review identifies “Remote Sensing” as the primary platform for FDDL research dissemination and highlights the field’s collaborative nature, with an average of 5.02 authors per paper. The co-occurrence network analysis reveals diverse research themes, spanning technical approaches and practical applications, with significant contributions from China, the United States, South Korea, Brazil, and Australia. Highly cited papers are explored, revealing their substantial influence on the field’s research focus. The analysis underscores the practical implications of integrating high-quality input data and advanced deep-learning techniques with remote sensing for effective fire detection. It provides actionable recommendations for future research, emphasizing interdisciplinary and international collaboration to propel FDDL technologies and applications. The study’s conclusions highlight the growing significance of FDDL technologies and the necessity for ongoing advancements in computational and remote sensing methodologies. The practical takeaway is clear: future research should prioritize enhancing the synergy between deep learning techniques and remote sensing technologies to develop more efficient and accurate fire detection systems, ultimately fostering groundbreaking innovations.
In recent years, a worldwide expansion in the frequency of large, uncontrolled, and catastrophic wildfire events has occurred, creating drastic social, economic, and environmental damage, especially in wildland–urban interface (WUI) zones. This damage includes losses of life, infrastructure, and ecosystem services. The impacts of wildfires at the WUI derive from the complicated and multidimensional interconnected relationships present in the Anthropocene. To enforce resilience of the environment and human communities against wildfires, it is critical to comprehend the local social-ecological systems holistically. In this paper, we present a theoretical framework approach, built on performance-based wildfire engineering, that is envisioned to be a stepping stone towards WUI resilience. To attain this objective, performance benchmarking and design is disaggregated into explicit components of a rigorous mathematical framework. They are linked to a causal inference chain, providing an integrated picture and enabling decision analysis to identify the optimal management strategies based on quantitative parameters. The proposed framework is developed from the total probability theorem and divides the risk assessment into single parts, in particular (1) hazard (wildfire) analysis, (2) social-ecological impact characterization, (3) social-ecological interaction analysis, (4) social-ecological impact analysis, (5) damage analysis, and (6) loss analysis. Therefore, the proposed framework can be applied by emergency agencies directly to assess the performance of society and ecosystem recovery after a wildfire, making emergency management and resilience policy-making more effective.
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 (i) fuels characterization, fire detection, and mapping; (ii) fire weather and climate change; (iii) fire occurrence, susceptibility, and risk; (iv) fire behavior prediction; (v) fire effects; and (vi) 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, to the end of 2019, we identified 300 relevant publications in which 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 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 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.
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.
Modeling the nonlinear and complex relationships between climate and tree-ring growth is of significance in dendroclimatic studies, but difficult to implement using traditional linear regression approaches. To overcome this difficulty, the technique of Artificial Neural Network (ANN) was employed in this study to develop the growth response models using the climate/growth database for Douglas-fir (Pseudotsuga menziesii var. menziesii [Mirb.] Franco) on southern Vancouver Island, Canada. The results show that the ANN models are able to extract nonlinear growth response patterns from the observed climate/tree-ring datasets, and to generate more accurate predictions than multiple linear regression approaches. The ANN-extracted climate-growth relationships can be displayed by scenario analysis; for example, when all other input variables are held fixed at their means, the limiting effect of April-July precipitation on tree growth decreases with increased precipitation. The main difficulty of applying ANN technique in dendroclimatology is the problem of overlearning (i.e., the ANN learns too many specific climate-growth patterns and loses the ability to generalize between similar climate-growth patterns). This problem can be alleviated by carefully designing the ANN, such as reducing the number of input variables, choosing a variety of training/testing sets, designing partially connected architectures with a small number of neurons in the hidden layer, and using early stopping during training process. The reliability of the derived ANN models is assessed by validation on independent testing datasets. The main advantages of the ANN technique over traditional dendroclimatic approaches are its ability to capture nonlinear climate-growth response, and its nonreliance on preassumed functional relationships for describing the observed datasets. The ANN method introduced in this article is sufficiently general to be applicable to many forest ecological modeling applications.
Human-caused forest fires are a serious problem throughout the world. Believing that there are predictable characteristics common to all fires, we analyzed the historical human-caused fire occurrence data for the Whitecourt Provincial Forest of Alberta using artificial neural network and geographic information system (ARC/INFO) technology. These data were also analyzed using logistic regression analysis (the binary logit model), which served as the 'domain expert' to identify the important input variables. A 314 fire and no-fire data set for the period 1986-1990 was used for training. The observations were whether at least one fire occurred, on a certain day, in one of the eight geographic zones defined within the study area. The models developed were tested using data from the 1991-1992 fire seasons, which had 58 fire observations. Using as input variables the Canadian Fire Weather Index for the day, area in km 2 of the geographic zone, and district (a 0/1 dummy variable from the logistic regression model, which accounts for observations within a forest district where human use is higher), the resultant model had four input nodes and two output nodes, and correctly predicted 85% of the no-fire observations and 78% of the fire observations.
Tree and shrub species composition and vegetation structure are key components influencing the quality of woodland or forest habitat for a wide range of organisms. This paper investigates the unique thematic classes that can be derived using integrated airborne LiDAR and spectral data. The study area consists of a heterogeneous, semi‐natural broadleaf woodland on an ancient site and homogeneous broadleaf and conifer woodland on an adjoining plantation. A parcel‐based unsupervised classification approach was employed, using the first two Principal Components from 12 selected wavebands of HyMap data and a Digital Canopy Height Model extracted from LiDAR data. The resultant 52 data clusters were amalgamated into 10 distinct thematic classes that contain information on species composition and vegetation structure. The thematic classes are relevant to the National Vegetation Classification (NVC) scheme for woodlands and scrub of Great Britain. Furthermore, in distinguishing structural subdivisions within the species‐based NVC classes, the thematic classification provides greater information for quantifying woodland habitat. The classes show degeneration from and regeneration to mature woodland communities and thus reflect the underlying processes of vegetation succession and woodland management. This thematic classification is ecologically relevant and is a forward development in woodland maps created from remote sensing data.
Fire occurrence in Mediterranean landscapes has been studied widely. Despite this, a specific monitoring of vegetation recovery after recurrent fires by means of satellite images has been developed to a lesser extent. With the use of Satellite Remote Sensing (SRS) techniques and multi-temporal Landsat images of the area of Ayora (287 700 ha) in Valencia (Eastern Spain), between the years 1984 and 1999, we studied the post-fire regeneration of the Normalized Difference Vegetation Index (NDVI) in areas subjected to different fire recurrences. Emphasis is given to the effect of time since fire, precipitation, and bedrock types on post-fire NDVI changes. Results suggest that for the first 7 years after a single fire, NDVI depends mainly on the time since fire (post-fire regeneration), whereas environmental parameters (precipitation and bedrock type) are of little relevance. After this period, precipitation begins to have a direct influence on the NDVI. In patches burned twice, with fire intervals of 8 and 9 years, NDVI is also controlled by the time since fire. Furthermore, NDVI recovery is faster after the first fire than after the second fire, suggesting that fire recurrence has a negative impact on the resilience of these communities. Bedrock type did not show any effect on NDVI after fire. These findings contribute to the understanding of Mediterranean landscape dynamics and provide evidence for the usefulness of NDVI in post-fire regeneration assessment, and the possible negative effects of the increasing fire recurrences observed in the last decades.
For estimating forest biomass and mapping of forest carbon content, an accurate classification of tropical forest stages is an important step. The successional stage of regenerating forests in a forest age map was a viable surrogate for forest biomass in an area north of Manaus City, Brazil. The forest stages considered were: (i) mature forest, (ii) regenerating forest <3 years, (iii) regenerating forest 3-5 years, (iv) regenerating forest 6-10 years and (v) regenerating forest 11-18 years. Areas of pasture were mapped also. Synthetic Aperture Radar (SAR) bands and optical TM (Thematic Mapper) bands were used to classify those classes, using the forest age map as the reference. The remotely sensed data comprised 20 bands (SAR bands from JERS-1, SIR-C and XSAR and optical bands from Landsat/TM) upon which discriminant analysis (DA) were used. DA results pointed to increased class discrimination when using SAR and TM data in relation to SAR data only. The bands selected were used as input to a neural network based classifier. Classification accuracy using SAR bands alone was around 30% for the six land cover classes. When regenerating forest stage classes were merged into a single class, the classification accuracy increased to around 80%. SAR data alone was unable to discriminate regenerating forest stages, having limited ability to discriminate between the subtle tonal/textural characteristics of each stage. A data set comprising TM and SAR bands showed increased classification accuracy in relation to SAR data alone, although some confusion between regenerating forest stages was still present. Following merging of regenerating forest stages into young (0-5 years) and intermediate (6-18 years), the overall accuracy was around 87%. The combination of SAR and TM bands were essential for the discrimination between regenerating forest stages. Pasture and mature forest were discriminated accurately in both SAR data alone and in the combined SAR and TM data.
Post-fire regeneration of Pinus halepensis, the most abundant tree in the Mediterranean Basin, can vary largely. In the present work, we aim to study the parameters determining this variability in the eastern Iberian Peninsula. For this reason we sampled in 2002 the sapling pine density on 22 plots that burned in 1993 and 71 plots that burned in 1994. Pre-fire vegetation (tree density and basal area) were obtained from the Spanish Forest Inventory. The regeneration ranged from 0.006 to 20.4pines/m2 (mean=1.24, S.D.=3.22). The statistical analysis suggested that the most important variables explaining this variability were the amount of branches found on the forest floor (branches collapsed from burned trees or branches left by foresters), the aspect of the plot, the pre-fire basal area, and whether the slope was terraced or not. High regeneration was observed in forests with large amounts of branches on the floor (which create appropriate microclimatic conditions), with northern aspects, with high pre-fire basal area, and on terraced slopes. Furthermore, other water-related variables (annual precipitation and slope) also had some (although lower) importance. These results have direct implications for forest managers in the study area.
A model based on an Artificial Neural Network (ANN) was developed for modeling and predicting red pine survival. The new model uses diameter at breast height and estimated annual diameter growth as its predictors. For training neural networks, a proportional coding scheme based on Gaussian distributions was used to transform the data into patterns of activities. Four model performance criteria--sum of square errors (SSE), χ² statistic, final predicted error (FPE), and predicted squared error (PSE)--were used to determine the adequacy of the new model. Based on the four criteria, the ANN-based new red pine survival model not only fits the data better than a statistical model; it is also expected to perform better on future data, provided that the training data are representative. The response surface of the ANN model shows it has the required flexibility to model red pine survival, especially in modeling both small and large but slow growing trees. This study also shows that a proportionally coded training data set may indeed be an effective form of input data representation for developing red pine survival models based on artificial neural networks. For. Sci. 37(5):1429-1440.
Reflectance-based indices derived from remote-sensing data have been widely used for detecting fire severity in forested areas. Rangeland ecosystems, such as sparsely vegetated shrub-steppe, have unique spectral reflectance dif-ferences before and after fire events that may not make reflectance-based indices appropriate for fire severity estimation. As an alternative, average vegetation height change (dh) derived from pre-and post-fire Light Detection and Ranging (LiDAR) data were used in this study for fire severity estimation. Theoretical deductions were conducted to demonstrate that LiDAR-derived dh is related to biomass combustion and thus can be used for fire severity estimation in rangeland areas. The Jeffreys–Matusita (JM) distance was calculated to evaluate the separability for each pair of fire severity classes, with an average JM distance of 1.14. Thresholds for classifying the level of fire severity were determined according to the mean and standard deviation of each class. A fire-severity classification map with 84% overall accuracy was obtained from the LiDAR dh method. Importantly, this method was sensitive to the difference between the moderate and high fire-severity classes.
Risk decision-making in natural hazards encompasses a plethora of environmental, socio-economic and management-related factors, and benefits greatly from exploring possible patterns and relations among these multivariate factors. Artificial neural networks, capable of general pattern classifications, are potentially well suited for risk decision support in natural hazards. This paper reports an example that assesses the risk patterns or probabilities of house survival from bushfires using artificial neural networks, with a simulation data set based on the empirical study by Wilson and Ferguson (Predicting the probability of house survival during bushfires, Journal of Environmental Management 23 (1986) 259–270). The aim of this study was to re-model and predict the relationship between risk patterns of house survival and a series of independent variables. Various configurations for input and output variables were tested using neural networks. An approach for converting linguistic terms into crisp numbers was used to incorporate linguistic variables into the quantitative neural network analysis. After a series of tests, results show that neural networks are capable of predicting risk patterns under all tested configurations of input and output variables, with a great deal of flexibility. Risk-based mathematical functions, be they linear or non-linear, can be re-modelled using neural networks. Finally, the paper concludes that the artificial neural networks serve as a promising risk decision support tool in natural hazards.
Artificial neural networks (ANNs) show a significant ability to discover patterns in data that are too obscure to go through
standard statistical methods. Data of natural phenomena usually exhibit significantly unpredictable non-linearity, but the
robust behavior of a neural network makes it perfectly adaptable to environmental models such as a wildland fire danger rating
system. These systems have been adopted by many developed countries that have invested in wildland fire prevention, and thus
civil protection agencies are able to identify areas with high probabilities of fire ignition and resort to necessary actions.
Since one of the drawbacks of ANNs is the interpretation of the final model in terms of the importance of variables, this
article presents the results of sensitivity analysis performed in a back-propagation neural network (BPN) to distinguish the
influence of each variable in a fire ignition risk scheme developed for Lesvos Island in Greece. Four different methods were
utilized to evaluate the three fire danger indices developed within the above scheme; three of the methods are based on network’s
weights after the training procedure (i.e., the percentage of influence—PI, the weight product—WP, and the partial derivatives—PD
methods), and one is based on the logistic regression (LR) model between BPN inputs and observed outputs. Results showed that
the occurrence of rainfall, the 10-h fuel moisture content, and the month of the year parameter are the most significant variables
of the Fire Weather, Fire Hazard, and Fire Risk Indices, respectively. Relative humidity, elevation, and day of the week have
a small contribution to fire ignitions in the study area. The PD method showed the best performance in ranking variables’
importance, while performance of the rest of the methods was influenced by the number of input parameters and the magnitude
of their importance. The results can be used by local forest managers and other decision makers dealing with wildland fires
to take the appropriate preventive measures by emphasizing on the important factors of fire occurrence.
The main study objective was to develop robust processing and analysis techniques to facilitate the use of small-footprint lidar data for estimating plot-level tree height by measuring individual trees identifiable on the three-dimensional lidar surface. Lidar processing techniques
included data fusion with multispectral optical data and local filtering with both square and circular windows of variable size. The lidar system used for this study produced an average footprint of 0.65 m and an average distance between laser shots of 0.7 m. The lidar data set was acquired
over deciduous and coniferous stands with settings typical of the southeastern United States. The lidar-derived tree measurements were used with regression models and cross-validation to estimate tree height on 0.017-ha plots. For the pine plots, lidar measurements explained 97 percent of
the variance associated with the mean height of dominant trees. For deciduous plots, regression models explained 79 percent of the mean height variance for dominant trees. Filtering for local maximum with circular windows gave better fitting models for pines, while for deciduous trees, filtering
with square windows provided a slightly better model fit. Using lidar and optical data fusion to differentiate between forest types provided better results for estimating average plot height for pines. Estimating tree height for deciduous plots gave superior results without calibrating the
search window size based on forest type.
Fire researchers and resource managers are dependent upon accurate, spatially-explicit forest structure information to support the application of forest fire behavior models. In particular, reliable estimates of several critical forest canopy structure metrics, including canopy bulk density, canopy height, canopy fuel weight, and canopy base height, are required to accurately map the spatial distribution of canopy fuels and model fire behavior over the landscape. The use of airborne laser scanning (LIDAR), a high-resolution active remote sensing technology, provides for accurate and efficient measurement of three-dimensional forest structure over extensive areas. In this study, regression analysis was used to develop predictive models relating a variety of LIDAR-based metrics to the canopy fuel parameters estimated from inventory data collected at plots established within stands of varying condition within Capitol State Forest, in western Washington State. Strong relationships between LIDAR-derived metrics and field-based fuel estimates were found for all parameters [sqrt(crown fuel weight): R2=0.86; ln(crown bulk density): R2=0.84; canopy base height: R2=0.77; canopy height: R2=0.98]. A cross-validation procedure was used to assess the reliability of these models. LIDAR-based fuel prediction models can be used to develop maps of critical canopy fuel parameters over forest areas in the Pacific Northwest.
The use of lidar data to estimate critical variables needed for modeling wildfire behavior was tested on a Scots pine forest (Pinus sylvestris L.) in central Spain. Lidar data accurately estimated crown bulk density at the plot level (r2=0.80). Lidar data could be used to directly estimate crown volume (r2=0.92) and foliage biomass (r2=0.84), which together produced better results than directly fitting the lidar data to crown bulk density. Incorporating equations that relate tree diameter at breast height and other forest parameters improved estimates of foliage biomass. Individual tree level analyses were not completely successful due to difficulty in accurately assigning laser pulses to the correct tree (r2=0.14).
The lack of maps depicting forest three-dimensional structure, particularly as pertaining to snags and understory shrub species distribution, is a major limitation for managing wildlife habitat in forests. Developing new techniques to remotely map snags and understory shrubs is therefore an important need. To address this, we first evaluated the use of LiDAR data for mapping the presence/absence of understory shrub species and different snag diameter classes important for birds (i.e. ≥ 15 cm, ≥ 25 cm and ≥ 30 cm) in a 30,000 ha mixed-conifer forest in Northern Idaho (USA). We used forest inventory plots, LiDAR-derived metrics, and the Random Forest algorithm to achieve classification accuracies of 83% for the understory shrubs and 86% to 88% for the different snag diameter classes. Second, we evaluated the use of LiDAR data for mapping wildlife habitat suitability using four avian species (one flycatcher and three woodpeckers) as case studies. For this, we integrated LiDAR-derived products of forest structure with available models of habitat suitability to derive a variety of species-habitat associations (and therefore habitat suitability patterns) across the study area. We found that the value of LiDAR resided in the ability to quantify 1) ecological variables that are known to influence the distribution of understory vegetation and snags, such as canopy cover, topography, and forest succession, and 2) direct structural metrics that indicate or suggest the presence of shrubs and snags, such as the percent of vegetation returns in the lower strata of the canopy (for the shrubs) and the vertical heterogeneity of the forest canopy (for the snags). When applied to wildlife habitat assessment, these new LiDAR-based maps refined habitat predictions in ways not previously attainable using other remote sensing technologies. This study highlights new value of LiDAR in characterizing key forest structure components important for wildlife, and warrants further applications to other forested environments and wildlife species.
This study compared two alternative techniques for predicting forest cover types from cartographic variables. The study evaluated four wilderness areas in the Roosevelt National Forest, located in the Front Range of northern Colorado. Cover type data came from US Forest Service inventory information, while the cartographic variables used to predict cover type consisted of elevation, aspect, and other information derived from standard digital spatial data processed in a geographic information system (GIS). The results of the comparison indicated that a feedforward artificial neural network model more accurately predicted forest cover type than did a traditional statistical model based on Gaussian discriminant analysis.
Two predictive modelling principles are discussed: multiple regression (MR) and neural networks (NN). The MR principle of linear modelling often gives low performance when relationships between variables are nonlinear; this is often the case in ecology; some variables must therefore be transformed. Despite these manipulations, the results often remain disappointing: poor prediction, dependence of residuals on the variable to predict. On the other hand NN are nonlinear type models. They do not necessitate transformation of variables and can give better results. The application of these two techniques to a set of ecological data (study of the relationship between density of brown trout spawning sites (redds) and habitat characteristics), shows that NN are clearly more performant than MR (R2 = 0.96 vs R2 = 0.47 or R2 = 0.72 in raw variables or nonlinear transformed variables). With the calculation power now currently available, NN are easy to implement and can thus be recommended for modelling of a number ecological processes.
Feed‐forward neural networks are now widely used in classification problems, whereas non‐linear methods of discrimination developed in the statistical field are much less widely known. A general framework for classification is set up within which methods from statistics, neural networks, pattern recognition and machine learning can be compared. Neural networks emerge as one of a class of flexible non‐linear regression methods which can be used to classify via regression. Many interesting issues remain, including parameter estimation, the assessment of the classifiers and in algorithm development.
Currently, commercial forestry applications of airborne scanning lidar are limited to geo-technical applications such as creation of digital terrain models for layout of roads or logging systems. We investigated the feasibility of predicting characteristics of forest stands with lidar data in a university-industry partnership. Lidar lends itself well to such applications because it allows direct measurement of important structural characteristics of height and canopy closure. We found that lidar data can be used to predict the stand characteristics of height, basal area, and volume quite well. The potential for commercial applications appears bright. Lidar data can be used to estimate stand characteristics over large areas or entire forests. After the process is streamlined, it should be possible to provide maps of height, basal area, and volume in such areas within a few weeks of the lidar collection flight.
This study relates the spatial variability in tree regeneration generated by fire with topography and pre-fire vegetation composition, and analyses how the pattern of fire severities determines post-fire regeneration of the dominant tree species (Pinus halepensis and Quercus ilex) in a large fire that occurred in north-eastern Spain in summer 1994. At the fire level, the proportion of the different fire severities in the burned area was characteristic of large fires that burn with high severity. At the level of plot, the variability of fire severity in the study site depended on two topographic characteristics: elevation and aspect. Plots burned with high fire severity were distributed at higher altitude than less severely burned plots, which were mainly distributed in south- and east-facing slopes. Fire severity also increased with Q. ilex density in the stand. The mosaic of fire severities determined both plant mortality and seedling regeneration. Mortality of stems caused by fire was very high in both species, but many Q. ilex individuals resprouted after fire. Seedling regeneration of P. halepensis showed large differences among plots in the burned area, reflecting large spatial variability due to the elevation gradient, the variation in density of adult pine trees, and the spatial variability created by fire.
Wildfires burn several million hectares in the United States annually. Time is critical in gathering information from burned landscapes for post-fire recovery planning. A technology to obtain spatial vegetation information across landscapes is Light Detecting and Ranging (LiDAR). We compared tree positional and height measurements, primarily from Douglas-fir (Pseudotsuga menziesii) and ponderosa pine (Pinus ponderosa), between field-based and LiDAR-derived measurements at three south-western Oregon (USA) sites. The sites represented a range of tree mortality from minimal to extensive. Our primary objective was to determine whether significant differences existed between field and LiDAR tree measurements in burned landscapes. Secondary objectives were to examine whether LiDAR pulse intensities in burned landscapes could differentiate coniferous from deciduous trees, discern fire-killed from live trees, and whether other tree measurement parameters were related to pulse intensities. No significant differences were detected between field-based and LiDAR-derived horizontal positions. Tree height differences between field-based and LiDAR measurements were significant at one site likely owing to dense canopy and measurement biases. Mean and maximum LiDAR intensities were significantly different between live and dead (fire-killed) trees in two of three sites. Additionally, crown diameter and tree sweep were significant in explaining variation in maximum LiDAR intensities at all sites.
This paper reports regeneration status of trees in two watersheds namely Phakot and Pathri Rao in Uttara- khand. Seedling, sapling and tree density were greater in Phakot watershed forest than those in Pathri Rao watershed forest. In general, both forests were regen- erating, although seedling and sapling population was higher in Phakot watershed forest. As far as the re- generation status is concerned, maximum tree species was found with fair regeneration in the forests of both the watersheds. In Phakot watershed, three species (Accacia nilotica, Engelhardtia spicata and Olea glan- dulifera) and in Pathri Rao watershed seven species (Acacia nilotica, Anogeissus latifolius, Casearia ellip- tica, Cassia fistula, Holarrhena pubescens, Mallotus phillippensis and Ougeinia oojeinensis) were found not regenerating. In Phakot watershed, general densities- diameters class distribution showed decline in density from small diameter class to higher diameter class whereas in Pathri Rao watershed no trend was evident.
The mean and standard deviation (SD) of light detection and ranging (LiDAR)-derived canopy height are related to forest structure. However, LiDAR data typically cover a limited area and have a high economic cost compared with satellite optical imagery. Optical images may be required to extrapolate LiDAR height measurements across a broad landscape. Different spectral indices were obtained from three Landsat scenes. The mean, median, SD and coefficient of variation (CV) of LiDAR canopy height measurements were calculated in 30-m square blocks corresponding with Landsat Enhanced Thematic Mapper Plus (ETM+) pixels. Correlation and forward stepwise regression analysis was applied to these data sets. Mean and median LiDAR height versus normalized difference vegetation index (NDVI), normalized difference moisture index (NDMI), normalized burn ratio (NBR) and wetness Tasseled Cap showed the best correlation coefficients (R ranging between −0.62 and −0.76). Nineteen regression models were obtained (R = 0.65–0.70). These results show that LiDAR-derived canopy height may be associated with Landsat spectral indices. This approach is of interest in sustainable forest management, although further research is required to improve accuracy.
Mean tree height, dominant height, mean diameter, stem number, basal area and timber volume of 116 georeferenced field sample plots were estimated from various canopy height and canopy density metrics derived by means of a small-footprint laser scanner over young and mature forest stands using regression analysis. The sample plots were distributed systematically throughout a 6500 ha study area, and the size of each plot was 232.9 m 2 . Regressions for coniferous forest explained 60-97% of the variability in ground reference values of the six studied characteristics. A proposed practical two-phase procedure for prediction of corresponding characteristics of entire forest stands was tested. Fifty-seven test plots within the study area with a size of approximately 3740 m 2 each were divided into 232.9 m 2 regular grid cells. The six examined characteristics were predicted for each grid cell from the corresponding laser data using the estimated regression equations. Average values for each test plot were computed and compared with ground-based estimates measured over the entire plot. The bias and standard deviations of the differences between predicted and ground reference values (in parentheses) of mean height, dominant height, mean diameter, stem number, basal area and volume were −0.58 to −0.85 m (0.64-1.01 m), −0.60 to −0.99 m (0.67-0.84 m), 0.15-0.74 cm (1.33-2.42 cm), 34-108 ha −1 (97-466 ha −1 ), 0.43-2.51 m 2 ha −1 (1.83-3.94 m 2 ha −1 ) and 5.9-16.1 m 3 ha −1 (15.1-35.1 m 3 ha −1 ), respectively.
To determine the effect of burnt tree removal on post-fire natural regeneration of Pinus halepensis, two 2 500m2 areas were selected six months after the fire in a totally destroyed mature (>70 years) pine forest. In one area, all the trees were cut down and removed 10 months after the fire and, in the other, all the trees were left standing (control). In each area, 20 permanent plots of 20m2 each were randomly placed, and all seedlings emerging within them labeled by individual numbered plastic tags. Emergence, mortality, density and growth (height) of 6649 P. halepensis seedlings were monitored during the first four post-fire years. Seedling emergence was concentrated in the first post-fire autumn–winter period. No positive effect on seedling emergence was detected as a consequence of burnt trunk dragging and subsequent turning over of soil. Wood removal produced an immediate average seedling mortality of 33%, and notably increased seedling mortality during the subsequent summer, probably due to increased exposure of seedlings to sunlight and the possible debilitation of many individuals by mechanical contact during burnt wood removal. A negative correlation of pine seedling mortality with height was detected, which increased significantly on wood removal in the third post-fire year. That is, short seedlings (
Feed-forward neural networks are now widely used in classification problems, whereas nonlinear methods of discrimination developed in the statistical field are much less widely known. A general framework for classification is set up within which methods from statistics, neural networks, pattern recognition and machine learning can be compared. Neural networks emerge as one of a class of flexible non-linear regression methods which can be used to classify via regression. Many interesting issues remain, including parameter estimation, the assessment of the classifiers and in algorithm development.
This research reports the major results from an evaluation of the first Nordic operational stand-based forest inventory using airborne laser scanner data. Laser data from a forest area of 250 km 2 were used to predict six biophysical stand variables used in forest planning. The predictions were based on regression equations estimated from 250 m 2 field training plots distributed systematically throughout the forest area. Test plots with an approximate size of 0.1-0.4 ha were used for validation. The testing revealed standard deviations between ground-truth values and predicted values of 0.36-1.37 m (1.9-7.6%) for mean height, 0.70-1.55 m (3.0-7.6%) for dominant height, 2.38-4.88 m 2 ha -1 (7.8-14.2%) for basal area and 13.9-45.9 m 3 ha -1 (6.5-13.4%) for stand volume. No serious bias was detected.
Artificial neural network (ANN) methods are used to predict forest characteristics. The data source is the Southeast Alaska (SEAK) Grid Inventory, a ground survey compiled by the USDA Forest Service at several thousand sites. The main objective of this article is to predict characteristics at unsurveyed locations between grid sites. A secondary objective is to evaluate the relative performance of different ANNs. Data from the grid sites are used to train six ANNs: multilayer perceptron, fuzzy ARTMAP, probabilistic, generalized regression, radial basis function, and learning vector quantization. A classification and regression tree method is used for comparison. Topographic variables are used to construct models: latitude and longitude coordinates, elevation, slope, and aspect. The models classify three forest characteristics: crown closure, species land cover, and tree size/structure. Models are constructed using n-fold cross-validation. Predictive accuracy is calculated using a method that accounts for the influence of misclassification as well as measuring correct classifications. The probabilistic and generalized regression networks are found to be the most accurate. The predictions of the ANN models are compared with a classification of the Tongass national forest in southeast Alaska based on the interpretation of satellite imagery and are found to be of similar accuracy. FOR. SCI. 50(2):259276.
We are concerned with feed-forward non-linear networks (multi-layer perceptrons, or MLPs) with multiple outputs. We wish to treat the outputs of the network as probabilities of alternatives (e.g. pattern classes), conditioned on the inputs. We look for appropriate output non-linearities and for appropriate criteria for adaptation of the parameters of the network (e.g. weights). We explain two modifications: probability scoring, which is an alternative to squared error minimisation, and a normalised exponential (softmax) multi-input generalisation of the logistic non-linearity. The two modifications together result in quite simple arithmetic, and hardware implementation is not difficult either. The use of radial units (squared distance instead of dot product) immediately before the softmax output stage produces a network which computes posterior distributions over class labels based on an assumption of Gaussian within-class distributions. However the training, which uses cross-class information, can result in better performance at class discrimination than the usual within-class training method, unless the within-class distribution assumptions are actually correct.
Quercus ilex and Pinus halepensis are two of the most common tree species of the western Mediterranean basin. Both species regenerate reliably after fire:
P. halepensis colonizes recently disturbed areas by effective seedling recruitment, while Q. ilex resprouts vigorously after disturbances. For this reason, the natural regeneration of these species after fire should ensure
the re-establishment of a forest similar to that which existed before the fire. This study analyzes with a simple simulation
model whether or not the relative abundance of monospecific and mixed forests of these species in the landscape is altered
by fire. We also analyze the topographic factors and the forest structure before the fire that determine the changes in forest
composition after fire. This study has been carried out in a large fire that occurred in NE Spain. Overall, 33% of plots changed
to another community type, but this probability of change varied considerably among community types before the fire. Monospecific
forests of P. halepensis or Q. ilex had a high probability of remaining in their original composition after the fire, whereas the resilience of mixed forests
of these two species was quite low. Mixed forests changed for the most part to monospecific P. halepensis or Q. ilex forests. Analysis of several factors determining these changes indicated that only elevation as a significant topographical
variable. The effect of fire was to increase the altitudinal differentiation between the two species. P. halepensis forests that changed to mixed or Q. ilex forests were those of highest elevation, while the mixed and Q. ilex plots that changed to P. halepensis forests were those located at the lowest elevations. Concerning structural variables before fire, density of Q. ilex trees before the fire showed a much greater effect than P. halepensis density in determining the post-fire community. Finally, burn severity also influenced the changes observed. For both P. halepensis and Q. ilex forests, plots that changed to another forest type were mainly those that burned more severely. In the case of mixed forests,
even low fire severities involved high probabilities of change to monospecific forests.
The utilization of pine trees bark is receiving increasing attention. For management purposes, it is very useful to be able to estimate bark volume quantity of standing trees. This paper is concerned with the use of artificial neural network (ANN) models for accurate bark volume estimation of standing pine trees (Pinus brutia) as an alternative to regression models. The procedure that should be followed in the development of such models is outlined. Five different nonlinear regression models were fitted to data using the Levenberg–Marquardt optimization algorithm. Because of the bias detected on the assumptions and the high error values, three-layer feed-forward cascade correlation ANN models were developed to estimate bark volume. A comparative study of both nonlinear regression models and selected cascade correlation ANN model with the structure of 4-7-1/0.9982 indicated that the ANN model performed better than the best nonlinear regression model with error values 6.02% (root mean square error) and 13.3% (Furnival's index) of the mean bark volume, respectively. Paired t-test, 45° line test and absolute deviation percentage were also used for validation of the chosen models. The results clearly demonstrate the superiority of the ANN models to the regression models due to their ability to overcome the problems in forest data, such as nonlinear relationships, non-Gaussian distributions, outliers and noise in the data. The ANN technique introduced in this study is sufficiently general and has great potential to be applicable to many forest modeling applications providing a very useful tool as an alternative to traditional regression models.
Due to increased fuel loading as a result of fire suppression, land managers in the American west are in need of precise information about the fuels they manage, including canopy fuels. Canopy fuel metrics such as canopy height (CH), canopy base height (CBH), canopy bulk density (CBD) and available canopy fuel (ACF) are specific inputs for wildfire behavior models such as FARSITE and emission models such as FOFEM. With finer spatial resolution data, accurate quantification of these metrics with detailed spatial heterogeneity can be accomplished. Light Detection and Ranging (LiDAR) and color near-infrared imagery are active and passive systems, respectively, that have been utilized for measuring a range of forest structure characteristics at high resolution. The objective of this research was to determine which remote sensing dataset can estimate canopy fuels more accurately and whether a fusion of these datasets produces more accurate estimates. Regression models were developed for ponderosa pine (Pinus ponderosa) stand representative of eastern Washington State using field data collected in the Ahtanum State Forest and metrics derived from LiDAR and imagery. Strong relationships were found with LiDAR alone and LiDAR was found to increase canopy fuel accuracy compared to imagery. Fusing LiDAR with imagery and/or LiDAR intensity led to small increases in estimation accuracy over LiDAR alone. By improving the ability to estimate canopy fuels at higher spatial resolutions, spatially explicit fuel layers can be created and used in wildfire behavior and smoke emission models leading to more accurate estimations of crown fire risk and smoke related emissions.
Forest inventories support resource managers to define the extent, size distribution, and species composition of forested and non-forested lands, and after re-measurement, the changes in such resources [Armitage, I., 1998. Guidelines for the management of tropical forests 1. The production of wood. FAO forestry paper 135].In order to maintain their validity and reliability, forest inventories need to be periodically and continuously updated affording expensive maintenance costs. This makes the development of reliable and efficient estimation methods to integrate traditional and direct data update techniques a necessity.Thanks to their flexibility and adaptability, artificial neural networks (ANNs) constitute a valid approach, as an alternative to traditional statistical methods, for modelling complex long-lived dynamic biological ecosystems such as forests.Using the large dataset of the Forest Inventory of the Special Administrative Province of Trento, Italy, two neural models, short-term model (STM) and long-term model (LTM), have been developed to update the tree diameter-at-breast-height (dbh) distributions for managed alpine stands.The combined use of ST and LT models allows forest managers to predict the number of standing trees in each 5-cm diameter class from “20 cm” to “80 cm and more” included in standard inventory forms, on a simulation lapse of time of 8 up to 25 years, for several alpine species: spruce (Picea abies), silver-fir (Abies alba), scots pine (Pinus sylvestris), black pine (Pinus nigra), swiss pine (Pinus cembra), larch (Larix decidua), beech (Fagus sylvatica) and other broadleaved species.Both models showed a good generalization capability and effectively reproduced the development of dbh distribution over time.The ST model showed a higher predictive power if compared to the LT model due to the fact that modelling difficulty increases proportionally to the length of simulation intervals.Finally, a simulator software (MODERNA), based on the ST and LT neural models, was developed to be integrated in the standard data acquisition procedures of the Forest and Wildlife Service of the Province of Trento.
Obtaining quantitative information about the recovery of fire-affected ecosystems is of utmost importance from the management and decision-making point of view. Nowadays the concern about natural environment protection and recovery is much greater than in the past. However, the resources and tools available for its management are still not sufficient. Thus, attention and precision is needed when decisions must be taken. Quantitative estimates on how the vegetation is recovering after a fire can be of help for evaluating the necessity of human intervention on the fire-affected ecosystem, and their importance will grow as the problem of forest fires, climate change and desertification increases.
The artificial neural network technique to model wind damage to forests was examined. The network used in the investigation was a three-layered feed-forward neural network with a backpropagation training-algorithm using a momentum term and flat spot elimination. To yield insights into the performance of the network, a logistic regression model was fitted as a baseline. Two different types of models were set up and analyzed for both approaches. A dichotomous model that predicted the categories “damaged” versus “undamaged” for two different damage thresholds and a multinomial model that predicted the damage in four damage classes. The performance of the network and the logistic regression model was measured using the mean squared sensitivity error. The results of the dichotomous model demonstrate that a feed-forward network is able to better classify forests susceptible to wind damage than a logistic regression model, especially when the frequency of the undamaged and damaged forest stands differs significantly. This study also shows that the network has a higher capacity to identify damaged forest stands, compared to the logistic regression model applied in this investigation. With the specific dataset used in the present study, the proportion of damaged forest stands predicted by the network was between the observed proportion and the proportion predicted by the logistic regression model. The results of the multinomial models showed that both, the statistical model and the neural network were unable to classify all four damage classes but showed a dichotomous behavior in predicting the damage only in the two extreme damage classes. Possibilities to optimize the network performance by using different training algorithms or topologies and principal differences between the two models referring to their specific properties are discussed.
Street trees positively affect the everyday life of city inhabitants and therefore successful management of this resource is important. A tree inventory is an essential first step towards this end, but this activity can be complex, costly and must be optimized. In order to reduce time and effort required for data acquisition either when using traditional field work or airborne laser systems, a method is proposed to predict the value of essential tree morphological parameters with surrogate variables and artificial intelligence multilayer perceptron networks (MLPs). To evaluate MLPs, seven different models were tested on Acer platanoides L., Acer saccharinum L., Celtis occidentalis L., Fraxinus pennsylvanica Marsh., Gleditsia triacanthos L., Tilia cordata Mill., and Ulmus pumila L. data sets. Three models were intended to predict diameter at breast height (DBH), annual DBH increment and crown volume with a minimal number of input measurements to be extracted from aerial LIDAR (Light Detection and Ranging) data. The last four were associated with traditional ground inventory methods and aimed to predict height, crown volume and their respective annual increments using less labour-intensive variables. The prediction performance was assessed with the Pearson r correlation coefficient, computed between the measured and estimated output values for each of the cross-validation test files per tree species and model. By using carefully selected biotic and abiotic input parameters, the prediction performance of multilayer perceptron showed robustness and precision despite different age–class distribution per species, dissimilar species morphological characteristics, uneven distribution of species within defined urban ecological zones, and varied abiotic growth conditions. More precisely, prediction coefficients were greater than 70% for all models with very small probability levels except for two predictions were input data exhibited strongly non-Gaussian distributions. Overall, the average prediction for all scenarios was 91%. Considering these results, it was found that prediction of DBH, annual DBH increment and crown volume is possible with limited aerial LIDAR laser information. Moreover, it was established that traditional field work effort can be further reduced by predicting the value of unmeasured morphological parameters within acceptable levels of precision. These findings can have an impact on future urban tree inventories. Depending on the number of trees to be measured, municipal administrations have the choice to use either airborne or traditional data acquisition methods. In both cases, this research proposes optimized procedures that may reduce the overall inventory costs.
Within forest growth modeling LOGIT models are used to predict individual tree mortality. In this paper we present, Multi-Layer Perceptron, Learning Vector Quantization and Cascade Correlation networks as different formalisms for mortality predictions. The data set for parameterizing the LOGIT model and training the different neural network types comes from the Austrian National Forest Inventory. After training the different network types, we evaluate the resulting mortality predictions using an independent data set from the Litschau forest. The results indicate that Multi-Layer Perceptron with the learning algorithm resilient back-propagation and scaled conjugate gradient and Cascade Correlation with learning algorithm resilient back-propagation perform the best predictions. This suggests that neural networks are a viable alternative to the conventional LOGIT approach.
Environmental flow estimation in regulated rivers has become a major issue for watershed management in Mediterranean countries. There are many methodologies for environmental flow computation, but they usually require accurate hydrological long-term flow records, which sometimes are unavailable, and/or extensive field measurement campaigns, which can be very costly especially when the environmental flows must be determined at many locations in large basins. We analyzed the potential of neural network models for the estimation of environmental flow values in gauging sections and reaches under a natural flow regime in the watershed of the Ebro River, Spain, with a view to a future application in both ungauged and/or regime-altered sections. Non-linear multilayer feed-forward cascade-correlation neural networks were developed to model the relationships between known environmental flows (Qb calculated) and two sets of independent variables related to physical and hydrological watershed characteristics or to the general flow regime. Three models were found capable of good estimations of environmental flows, based on variables such as the 10-year average of the lower monthly flows, the mean value of the length of the period in days with flows continually below the 40% of the mean annual flow (spell duration), and the flow equaled or exceeded 270 days per year (Q270). Correlation coefficients (r) between calculated and estimated values were high (>0.90), and average absolute errors were low (<0.44 m3/s) for the three models. The limited number of variables in the models (just two) was considered very promising for operational application of the model to ungauged or regime-altered sections. Results suggest that artificial neural network models can be simple, robust, reliable and cost-efficient tools for environmental flow determination at the watershed level.
This paper rigorously establishes that standard multilayer feedforward networks with as few as one hidden layer using arbitrary squashing functions are capable of approximating any Borel measurable function from one finite dimensional space to another to any desired degree of accuracy, provided sufficiently many hidden units are available. In this sense, multilayer feedforward networks are a class of universal approximators.
Wild-land fires have become intense and more frequent all over the world. Improving the accuracy of mapping fuel models is essential for fuel management decisions and explicit fire behavior prediction for real-time support of suppression tactics and logistics decisions. The overall aim of this paper is to develop the use of lidar (LIght Detection and Ranging) remote sensing to accurately and effectively assess fuel models in East Texas. More specific goals include: (1) developing lidar derived products and the methodology to use them for assessing fuel models; (2) investigating the use of several techniques for data fusion of lidar and multispectral imagery for assessing fuel models; (3) investigating the gain in fuels mapping accuracy when using lidar as opposed to QuickBird imagery alone; and (4) producing spatially explicit digital fuel maps. Estimates of fuel models were compared with in-situ data collected over 62 plots. We employ a unique approach to classify fuel models using a combination of lidar height bins and multispectral image data. Different image processing approaches for fusing lidar and multispectral data, such as the Minimum Noise Fraction (MNF) and Principle Component Analysis (PCA), were used to improve the overall accuracy of image classification. Supervised image classification methods provided better accuracy (90.10%) with the fusion of airborne lidar data with QuickBird data than with QuickBird imagery alone (76.52%).According to our results, lidar derived data provide accurate estimates of surface fuel parameters efficiently and accurately over extensive areas of forests. This study demonstrates the importance of using accurate maps of fuel models derived using new lidar remote sensing techniques.
A feedforward artificial neural network, coupled with a regional GIS (geographic information system), is described that is being used to assess the potential impacts of climate change on a complex landscape of tropical forests. The model quantifies the relative suitability of environments for 15 forests classes using the best information that is available: a structural-environmental classification of forest types, vegetation maps and spatial estimates of environmental variables. Inputs to the model include climate variables, soil parent material classes and terrain variables. The model is highly successful at distinguishing the relative suitability of environments for the forest classes with 75% of the forest mosaic accurately predicted by the model at a one hectare resolution over more than two million hectares. The model was used to estimate potential forest distributions in several climates occurring since the end of the last glacial period. These distributions shift dramatically in response to scenarios representing past climates. Certain locations are occupied by a forest class in only some climates while others are always occupied by the same class despite large changes in regional mean annual temperature and precipitation. Using the model to assess the possible impacts of future climate change and estimating the pre-settlement distribution of forest types in the region is also discussed. The coupling of neural networks with a cellular automata model is also described as a means to assess the importance of spatial constraints on the potential redistribution of forest types in the future. The usefulness of artificial neural networks when applied to vegetation change studies in our region suggests that this approach could be applied in many tropical regions, where floristic diversity is high and mechanistic understanding is comparatively low.
The overall goal of this study was to develop methods for assessing crown base height for individual trees using airborne lidar data in forest settings typical for the southeastern United States. More specific objectives are to: (1) develop new lidar-derived features as multiband height bins and processing techniques for characterizing the vertical structure of individual tree crowns; (2) investigate several techniques for filtering and analyzing vertical profiles of individual trees to derive crown base height, such as Fourier and wavelet filtering, polynomial fit, and percentile analysis; (3) assess the accuracy of estimating crown base height for individual trees, and (4) investigate which type of lidar data, point frequency or intensity, provides the most accurate estimate of crown base height. A lidar software application, TreeVaW, was used to locate individual trees and to obtain per tree measurements of height and crown width. Tree locations were used with lidar height bins to derive the vertical structure of tree crowns and measurements of crown base height. Lidar-derived crown base heights of individual trees were compared to field observations for 117 trees, including 94 pines and 23 deciduous trees. Linear regression models were able to explain up to 80% of the variability associated with crown base height for individual trees. Fourier filtering used for smoothing the vertical crown profile consistently provided the best results when estimating crown base height.
Forest fire management practices are highly dependent on the proper monitoring of the spatial distribution of the natural and man-made fuel complexes at landscape level. Spatial patterns of fuel types as well as the three-dimensional structure and state of the vegetation are essential for the assessment and prediction of forest fire risk and fire behaviour. A combination of the two remote sensing systems, imaging spectrometry and light detection and ranging (LiDAR), is well suited to map fuel types and
properties, especially within the complex wildland–urban interface. LiDAR observations sample the spatial information dimension providing explicit geometric information about the structure of the Earth’s surface and super-imposed objects. Imaging spectrometry on the other hand samples the spectral dimension, which is sensitive for discrimination of surface types. As a non-parametric classifier support vector machines (SVM) are particularly well adapted to classify data of high dimensionality and from multiple sources as proposed in this work. The presented approach achieves an improved land cover
mapping adapted to forest firemanagement needs. The map is based on a single SVM classifier combining the spectral and spatial information dimensions provided by imaging spectrometry and LiDAR.
Artificial neural networks (ANNs) are non-linear mapping structures based on the function of the human brain. They have been shown to be universal and highly flexible function approximators for any data. These make powerful tools for models, especially when the underlying data relationships are unknown. In this reason, the international workshop on the applications of ANNs to ecological modelling was organized in Toulouse, France (December 1998). During this meeting, we discussed different methods, and their reliability to deal with ecological data. The special issue of this ecological modelling journal begins with the state-of-the-art with emphasis on the development of structural dynamic models presented by S.E. Jorgensen (DK). Then, to illustrate the ecological applications of ANNs, examples are drawn from several fields, e.g. terrestrial and aquatic ecosystems, remote sensing and evolutionary ecology. In this paper, we present some of the most important papers of the first workshop about ANNs in ecological modelling. We briefly introduce here two algorithms frequently used; (i) one supervised network, the backpropagation algorithm; and (ii) one unsupervised network, the Kohonen self-organizing mapping algorithm. The future development of ANNs is discussed in the present work. Several examples of modelling of ANNs in various areas of ecology are presented in this special issue.
Meaningful relationships between forest structure attributes measured in representative field plots on the ground and remotely sensed data measured comprehensively across the same forested landscape facilitate the production of maps of forest attributes such as basal area (BA) and tree density (TD). Because imputation methods can efficiently predict multiple response variables simultaneously, they may be usefully applied to map several structural attributes at the species-level. We compared several approaches for imputing the response variables BA and TD, aggregated at the plot-scale and species-level, from topographic and canopy structure predictor variables derived from discrete-return airborne LiDAR data. The predictor and response variables were associated using imputation techniques based on normalized and unnormalized Euclidean distance, Mahalanobis distance, Independent Component Analysis (ICA), Canonical Correlation Analysis (aka Most Similar Neighbor, or MSN), Canonical Correspondence Analysis (aka Gradient Nearest Neighbor, or GNN), and Random Forest (RF). To compare and evaluate these approaches, we computed a scaled Root Mean Square Distance (RMSD) between observed and imputed plot-level BA and TD for 11 conifer species sampled in north-central Idaho. We found that RF produced the best results overall, especially after reducing the number of response variables to the most important species in each plot with regard to BA and TD. We concluded that RF was the most robust and flexible among the imputation methods we tested. We also concluded that canopy structure and topographic metrics derived from LiDAR surveys can be very useful for species-level imputation.