ArticlePDF Available

Abstract and Figures

Reliable identification of clouds is necessary for any type of optical remote sensing image analysis, especially in operational and fully automatic setups. One of the most elaborated and widespread algorithms, namely Fmask, was initially developed for the Landsat suite of satellites. Despite their similarity, application to Sentinel-2 imagery is currently hampered by the unavailability of a thermal band, and although results can be improved when taking the cirrus band into account, Sentinel-2 cloud detections are unsatisfactory in two points. (1) Low altitude clouds can be undetectable in the cirrus band, and (2) bright land surfaces – especially built-up structures – are often misclassified as clouds when only considering spectral information. In this paper, we present the Cloud Displacement Index (CDI), which makes use of the three highly correlated near infrared bands that are observed with different view angles. Hence, elevated objects like clouds are observed under a parallax and can be reliably separated from bright ground objects. We compare CDI with the currently used cloud probabilities, and propose how to integrate this new functionality into the Fmask algorithm. We validate the approach using test images over metropolitan areas covering a wide variety of global environments and climates, indicating the successful separation of clouds and built-up structures (overall accuracy 95%, i.e. an improvement in overall accuracy of 0.29–0.39 compared to the previous Fmask versions over the 20 test sites), and hence a full compensation for a missing thermal band.
Content may be subject to copyright.
A preview of the PDF is not available
... For each time period, we applied the same processing and classification techniques. We first applied the F mask algorithm (Frantz et al., 2018) to remove cloud cover and then used the modified Otsu thresholding algorithm (Karuppanagounder and Genish, 2012) to separate land from water. As Wicaksono and Hafizt (2013) advise, we then applied GEE's 'hazeRemovalDeepwater' function for atmospheric correction using dark pixel subtraction to remove hazing. ...
... Google Earth Engine platform was used to access all images. The F mask algorithm was used to mask out clouds in Sentinel 2 images and Landsat 5 images as elaborated by Frantz et al. (2018). Lee speckle filter was used to correct the speckle noise in Sentinel 1 and ALOS PALSAR as elaborated by Lee et al. (1994). ...
... changes can also remove or at least reduce the BRDF differences embedded in the satellite data, and in this way, the change pixel can be correctly identified (Fig. 10b) using a time-series based change detection algorithm . It is worth noting that the angular information can be useful for identifying the target and location of land change, such as improving land cover classification (Jiao et al., 2011), detecting moving objects such as clouds (Frantz et al., 2018), aircraft (Liu et al., 2020b), and detection of newly built houses (Huang et al., 2020), due to the inclusion of 3D information. ...
Article
Full-text available
The discipline of land change science has been evolving rapidly in the past decades. Remote sensing played a major role in one of the essential components of land change science, which includes observation, monitoring, and characterization of land change. In this paper, we proposed a new framework of the multifaceted view of land change through the lens of remote sensing and recommended five facets of land change including change location, time, target, metric, and agent. We also evaluated the impacts of spatial, spectral, temporal, angular, and data-integration domains of the remotely sensed data on observing, monitoring, and characterization of different facets of land change, as well as discussed some of the current land change products. We recommend clarifying the specific land change facet being studied in remote sensing of land change, reporting multiple or all facets of land change in remote sensing products, shifting the focus from land cover change to specific change metric and agent, integrating social science data and multi-sensor datasets for a deeper and fuller understanding of land change, and recognizing limitations and weaknesses of remote sensing in land change studies.
... changes can also remove or at least reduce the BRDF differences embedded in the satellite data, and in this way, the change pixel can be correctly identified (Fig. 10b) using a time-series based change detection algorithm . It is worth noting that the angular information can be useful for identifying the target and location of land change, such as improving land cover classification (Jiao et al., 2011), detecting moving objects such as clouds (Frantz et al., 2018), aircraft (Liu et al., 2020b), and detection of newly built houses (Huang et al., 2020), due to the inclusion of 3D information. ...
... Despite this, the runoff derived from Sentinel-2 is not better correlated with the observational data than the estimates derived from Landsat. This finding may be because Sentinel-2 is not equipped with a thermal sensor, which limits its ability to detect low altitude clouds (Zhu, Wang and Woodcock 2015;Frantz et al. 2018). For example, inaccuracies of detection for a scene acquired by Sentinel-2 on 23 August 2018, cause the Nash, R, and RMSE values to change from 0.86, 0.92, and 3596.2 m 3 /s to 0.77, 0.89, and 4487.3 m 3 /s, respectively, for the entire Sentinel-2-derived time series. ...
Article
River discharge plays a vital role in the Earth’s water cycle. Recent studies have sought to estimate river discharge using optical sensors onboard satellites, typically deriving discharge from multispectral data after excluding many images that are contaminated by cloud or affected by the scan-line-corrector (SLC) failure on Landsat 7. As such, time series of estimated discharge generally include a high proportion of missing values. To overcome these issues, we propose an easy and feasible methodology to temporally densify the time series of discharge estimates by repairing cloud-contaminated and SLC-affected images on the Google Earth Engine platform. The middle Yangtze River (MYR) was chosen as the study area to demonstrate and evaluate the performance of the methodology. First, different water indices were compared to select the most appropriate indicator for water surface retrieval along the MYR. The Pekel method shows the best performance for extracting the water surface when turbid or under thin, semi-transparent clouds. Second, the performances of the Tourian method and the Manning equation for estimating the discharge were compared when using only good-quality images, yielding Nash-Sutcliffe efficiency (Nash) coefficients of 0.69 and 0.79, respectively. Images affected by the SLC and over 50% cloud cover accounted for 41% of 532 records used from 1986 to 2019. The repaired, temporally complete time series of discharge estimates by the Manning equation corresponds to a correlation coefficient of 0.86, a Nash coefficient of 0.73, and a root mean square error of 5697 m³/s from 2000 to 2019. These results show the strong potential of this method for estimating the discharge of rivers with known cross-sectional and the morphological conditions worldwide.
Article
Accurate detection of cloud cover is an important task in the field of Remote Sensing of the environment. Currently, a lot of development is going on in this field by using various methods. Some of the methods apply concepts of machine learning (ML) whereas some apply deep learning. Since the accuracy of ML being lower than deep learning, the latter is preferred. This paper also uses the method of deep learning to detect cloud cover using images of satellite. This paper proposes a modified U-Net based deep learning model for cloud cover detection in satellite images. The model proposed is not as accurate as the original model, but it compensates for it by reducing the time for learning. The accuracy of the model came out to be 89.73%.
Preprint
Full-text available
Most readily available landuse/landcover (LULC) data are developed using growing season remote sensing images and/or annual time steps. We used new Dynamic World near real-time global LULC to compare how geospatial environmental models of water quality and hydrology respond to growing vs. non-growing season LULC data. Non-growing season LULC had more built area and less tree cover than growing season data due to seasonal impacts on classifications. We evaluated the impacts of these seasonal LULC estimate differences on water quality and quantity models that span a range of complexity, including the Soil and Water Assessment Tool (SWAT). We found that in mixed-LULC watersheds, seasonal LULC classification differences could cause large differences in model outputs depending on the LULC season used. Within reason, model parameter optimization may compensate for these differences using separate calibration for each season. These findings provide opportunities for further investigations with hydrologic, climate, biogeochemical, and ecological models.
Article
Vegetation restoration potential (VRP) mapping provides important information for ecosystem restoration planning. However, the inappropriate assumption traditional models that VRPs are identical within an individual similar habitat unit may result in low accuracy of VRP maps. This study proposes an improved data‐driven model, namely, the similar habitat and machine learning based VRP mapping (SHMLVRPM) model. This new model introduces a variety of machine learning models to mine information on geographical environment heterogeneity in areas of similar habitat, which helps to improve the accuracy of VRP maps. Taking Yan’an City, Shanxi Province, China as our study area, we demonstrate the modeling process and validate the model. Our results show that the SHMLVRPM model can effectively construct high‐accuracy vegetation restoration potential maps, and its information entropy is approximately 5.8 greater than that derived from the traditional model. The random forest method has the highest prediction accuracy (R2=0.8) among the tested machine learning methods. The average VRP value of Yan’an is approximately 68%; counties with the low VRP achievement are concentrated in the northern part of Yan’an, only 54%. Our research results can assist policymakers in optimizing vegetation restoration options and promoting the protection and sustainable development of fragile ecosystems.
Article
Long believed to be insignificant, melt activity on the Northeast Greenland Ice Stream (NEGIS) has increased in recent years. Summertime Arctic clouds have the potential to strongly affect surface melt processes by regulating the amount of radiation received at the surface. However, the cloud effect over Greenland is spatially and temporally variable and high‐resolution information on the northeast is absent. This study aims at exploring the potential of a high‐resolution configuration of the polar‐optimized Weather Research & Forecasting (PWRF) model in simulating cloud properties in the area of the Nioghalvfjerdsfjorden Glacier (79N). Subsequently, the model‐simulations are employed to investigate the impact of Arctic clouds on the surface energy budget and on surface melting during the extensive melt event at the end of July 2019. Compared to Automatic Weather Station (AWS) measurements and remote sensing data (Sentinel‐2A and the Moderate Resolution Imaging Spectroradiometer, MODIS), PWRF simulates cloud properties with sufficient accuracy. It appears that peak melt was caused by an increase in solar radiation and sensible heat flux (SHF) in response to a blocking anticyclone and foehn winds under absence of clouds. Cloud warming over high‐albedo surfaces helped to precondition the surface and prolonged the melting as the anticyclone abated. The results are sensitive to the surface albedo and suggest spatiotemporal differences in the cloud effect as snow and ice properties change over the course of the melting season. This demonstrates the importance of including high‐resolution information on clouds in analyses of ice sheet dynamics.
Article
Dynamic monitoring of floods is important for water resource management and disaster prevention. Obtaining multitemporal surface water distribution maps using remote sensing technology can help in elucidating the trends in water expansion so that measures can be quickly formulated. Sentinel-1 synthetic aperture radar (SAR) observation data are particularly suitable for this task because of their high spatial resolution and short revisit cycle, as well as its cloud-penetration ability. However, quickly and accurately mapping floods from a large number of SAR images remains challenging because of the enormous pressure on data acquisition and processing. Hence, in this study, we designed a new automatic SAR image flood mapping method based on the Google Earth Engine (GEE) cloud platform, which is an improvement over the Otsu method, and solves the problem of a higher segmentation threshold caused by images that do not meet the bimodal distribution hypothesis. In addition, to eliminate the omissions caused by salt-and-pepper noise and the misclassification caused mainly by low-backscattering-intensity vegetation and mountain shadows, we constructed an algorithm based on topological relationships and a DSM (Digital Surface Model) local search algorithm. The proposed method achieved an accuracy of 96.213% and 98.611% and F1 scores of 0.87254 and 0.89298 for plains and mountainous terrain, respectively. This method uses powerful computing resources and abundant datasets provided by the GEE cloud platform, and can be used for large-scale, long-term, and dynamic flood monitoring.
Article
Full-text available
Dry tropical forests undergo massive conversion and degradation processes. This also holds true for the extensive Miombo forests that cover large parts of Southern Africa. While the largest proportional area can be found in Angola, the country still struggles with food shortages, insufficient medical and educational supplies, as well as the ongoing reconstruction of infrastructure after 27 years of civil war. Especially in rural areas, the local population is therefore still heavily dependent on the consumption of natural resources, as well as subsistence agriculture. This leads, on one hand, to large areas of Miombo forests being converted for cultivation purposes, but on the other hand, to degradation processes due to the selective use of forest resources. While forest conversion in south-central rural Angola has already been quantitatively described, information about forest degradation is not yet available. This is due to the history of conflicts and the therewith connected research difficulties, as well as the remote location of this area. We apply an annual time series approach using Landsat data in south-central Angola not only to assess the current degradation status of the Miombo forests, but also to derive past developments reaching back to times of armed conflicts. We use the Disturbance Index based on tasseled cap transformation to exclude external influences like inter-annual variation of rainfall. Based on this time series, linear regression is calculated for forest areas unaffected by conversion, but also for the pre-conversion period of those areas that were used for cultivation purposes during the observation time. Metrics derived from linear regression are used to classify the study area according to their dominant modification processes. We compare our results to MODIS latent integral trends and to further products to derive information on underlying drivers. Around 13% of the Miombo forests are affected by degradation processes, especially along streets, in villages, and close to existing agriculture. However, areas in presumably remote and dense forest areas are also affected to a significant extent. A comparison with MODIS derived fire ignition data shows that they are most likely affected by recurring fires and less by selective timber extraction. We confirm that areas that are used for agriculture are more heavily disturbed by selective use beforehand than those that remain unaffected by conversion. The results can be substantiated by the MODIS latent integral trends and we also show that due to extent and location, the assessment of forest conversion is most likely not sufficient to provide good estimates for the loss of natural resources.
Article
Full-text available
Sentinel-2 is a very new programme of the European Space Agency (ESA) that is designed for fine spatial resolution global monitoring. Sentinel-2 images provide four 10 m bands and six 20 m bands. To provide more explicit spatial information, this paper aims to downscale the six 20 m bands to 10 m spatial resolution using the four directly observed 10 m bands. The outcome of this fusion task is the production of 10 Sentinel-2 bands with 10 m spatial resolution. This new fusion problem involves four fine spatial resolution bands, which is different to, and more complex than, the common pan-sharpening fusion problem which involves only one fine band. To address this, we extend the existing two main families of image fusion approaches (i.e., component substitution, CS, and multiresolution analysis, MRA) with two different schemes, a band synthesis scheme and a band selection scheme. Moreover, the recently developed area-to-point regression kriging (ATPRK) approach was also developed and applied for the Sentinel-2 fusion task. Using two Sentinel-2 datasets released online, the three types of approaches (eight CS and MRA-based approaches, and ATPRK) were compared comprehensively in terms of their accuracies to provide recommendations for the task of fusion of Sentinel-2 images. The downscaled ten-band 10 m Sentinel-2 datasets represent important and promising products for a wide range of applications in remote sensing. They also have potential for blending with the upcoming Sentinel-3 data for fine spatio-temporal resolution monitoring at the global scale.
Article
Tropical dry forests provide globally important ecosystem services and host exceptionally high biodiversity. These biomes are currently under immense pressure, particularly for conversion to agriculture, and already experience high global deforestation rates. Miombo forests in Southern Angola are affected by deforestation, fragmentation and degradation, caused mainly by an increasing rural population who follows a traditional farming system of shifting cultivation with slash-and-burn agriculture. After the termination of the civil war in 2002, population growth and resettlements have accelerated the use of woody resources, selective logging and clearing for cultivation purposes and led to an exceedance of sustainability thresholds. Large scale projects are expected to put further pressure on the forests and increase the potential of conflicts regarding land resources and competition with local subsistence farming.
Article
Land-use is transforming habitats across the globe, thereby threatening wildlife. Large mammals are especially affected because they require large tracts of intact habitat and functioning corridors between core habitat areas. Accurate land-cover data is critical to identify core habitat areas and corridors, and medium resolution sensors such as Landsat 8 provide opportunities to map land cover for conservation planning. Here, we used all available Landsat 8 imagery from launch through December 2014 to identify large mammal corridors and assess their quality across the Caucasus Mountains (> 700,000 km²). Specifically, we tested the usefulness of seasonal image composites (spring, summer, fall, and winter) and a range of image metrics (e.g., mean and median reflectance across all clear observations) to map nine land-cover classes with a Random Forest classifier. Using image composites from all four seasons yielded markedly higher overall accuracy than using single-season composites (8% increase) and the inclusion of image metrics further improved the classification significantly. Our final land-cover map had an overall accuracy of 85%. Using our land-cover map, we parameterized connectivity models for three generic large mammal groups and identified wildlife corridors and bottlenecks within corridors with cost-distance modeling and circuit theory. Corridors were numerous (in total, 85, 131, and 132 corridors for our three mammal groups, respectively), but often had bottlenecks or high average cost along the least-cost path, indicating limited functioning. Our findings highlight the potential of Landsat 8 composites to support connectivity analyses across large areas, and thus to contribute to conservation planning, and serve as an early warning system for biodiversity loss in areas where on-the-ground monitoring is challenging, such as in the Caucasus.
Article
The need for operational monitoring of landscape processes on the national to global scale led to an increased demand for pixel-based composites using complete earth observation (EO) archives. Commonly, composites are generated without explicit consideration of temporal criteria but are rather based on optimizing band indices within a pre-defined temporal window. However, for certain applications phenology-adapted composites that represent the land surface as being in the same phenological stage are required, e.g. tree type discrimination where greening up or senescence dates are modified by terrain elevation. We developed a novel pixel-based compositing technique that dynamically adjusts the selection process to the underlying land surface phenology (LSP) of each pixel. By doing so, phenologically sound composites across large areas can be derived for regular intervals and different phenological points in time, e.g. peak, end or minimum of season. Various day-of-year (DOY) scoring functions were implemented to flexibly define the phenological target. The technique is general enough for global application and can be applied to any kind of gridded EO archive, herein demonstrated for MODIS and Landsat data. Multi-annual composites were successfully generated for Zambia for most seasons. As an exception, we found even very frequent MODIS observations to be insufficient for peak vegetation composites due to interference with the rainy season. The phenology-adaptive composites were compared to static ones, i.e. using a single target DOY. Results clearly indicated that biomass levels differ significantly between the techniques, and a phenological normalization across elevation gradients and land cover classes could be achieved. However, the implications are non-trivial and the characteristics of both methods need to be considered cautiously before deciding which approach is superior with regards to a specific thematic application. The quality of the MODIS and Landsat composites, as well as the performance of the phenology-adaptive and static compositing techniques were assured using a quantitative cross-comparison. A 12-year annual time series demonstrated the feasibility for land cover change and modification mapping. Several change processes were clearly discriminable. The resulting phenologically coherent composites are important to establish national, regional or even global landscape monitoring, reporting and verification systems.
Article
As part of the Copernicus programme of the European Union (EU), the European Space Agency (ESA) has developed and is currently operating the Sentinel-2 mission that is acquiring high spatial resolution optical imagery. This paper provides a description of the calibration activities and the current status of the mission products validation activities. Measured performances, from the validation activities, cover both Top-Of-Atmosphere (TOA) and Bottom-Of-Atmosphere (BOA) products. Results presented in this paper show the good quality of the mission products both in terms of radiometry and geometry and provide an overview on next mission steps related to data quality aspects.
Article
The surface reflectance, i.e., satellite derived top of atmosphere (TOA) reflectance corrected for the temporally, spatially and spectrally varying scattering and absorbing effects of atmospheric gases and aerosols, is needed to monitor the land surface reliably. For this reason, the surface reflectance, and not TOA reflectance, is used to generate the greater majority of global land products, for example, from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) sensors. Even if atmospheric effects are minimized by sensor design, atmospheric effects are still challenging to correct. In particular, the strong impact of aerosols in the visible and near infrared spectral range can be difficult to correct, because they can be highly discrete in space and time (e.g., smoke plumes) and because of the complex scattering and absorbing properties of aerosols that vary spectrally and with aerosol size, shape, chemistry and density.