Burned area mapping using active and passive sensors of medium spatial resolution

Abstract

This doctoral thesis focuses on the development of new algorithms for burned area mapping using active and passive sensors of medium spatial resolution. Such algorithms shall allow reducing current uncertainties on areas affected by fire every year at global level. Currently, global burned area is estimated from low spatial resolution optical images, which have limitations in areas with high cloud cover, as well as when mapping small fires (i.e., less than 100 hectares). The developed algorithms reduce such limitations by using radar images and by combining radar and optical datasets acquired at medium spatial resolution by the European Space Agency (ESA) satellites Sentinel-1 and Sentinel-2. The thesis is structured in eight chapters. The first chapter describes the importance of fire at global and regional level, and highlighting the need for accurate burned area products for a range of application including ecosystem management, GHGs emissions monitoring and vegetation modelling. Since its inception, remote sensing has been a highly valuable source of information for burned area detection and monitoring. Therefore, the physical principles on which burned area mapping is achieved through using remote sensing data are also explained and related to the state-of-the-art when mapping burned areas through optical images, radar images, and their combination. Finally, the hypothesis, motivation, and objectives of the thesis are presented. In the second chapter, a burned area mapping algorithm based on C-band Sentinel-1 imagery is discussed. The algorithm applies the Reed-Xiaoli detector (RXD) to distinguish anomalous changes of the backscatter coefficient. Such changes are linked to fires events through spatially and temporally coincident thermal anomalies acquired by ancillary sensors. For periods with no thermal anomalies, a machine learning classifier, random forests, was used to detect the burned areas. Burned area perimeters derived from optical images (Landsat-8 and Sentinel-2) were used to validate the algorithm over 21 million hectares distributed worldwide in 18 locations that represent the main biomes affected by fires. The mean Dice coefficient (DC) was calculated over all the 18 locations revealed a burned area mapping accuracy of 0.59±0.06 ( confidence interval, 95%), with the mean errors of omission (OE) and commission (CE) reaching 0.43±0.08 and 0.37±0.06, respectively. The results were Compared with those provided by the most widely used global burned area product, the MCD64A1. The proposed algorithm improved, on average, the DC by 0.13 by reducing OE (0.12) and CE (0.06). In the third chapter, and the relationship between mapping accuracy and computing time needed for the RXD-based algorithm, described in chapter two, was analysed for different pixel spacings (i.e., 20, 30, 40 and 50 m). The analysis was carried out in six globally distributed study areas. The results suggest marginal differences in accuracy when varying the pixel spacing, with slightly more accurate maps being obtained using higher spacings when compared to the nominal of Sentinel-1 spatial resolution (20 m). However, the computing time was considerably higher at low pixel spacing with images between 30-50 m providing the optimum trade-off between accuracy and computing time. In the fourth chapter, factors that may influence the burned area mapping accuracy were analysed. The main focus of the analysis was the temporal decorrelation process, observed during algorithm development. The temporal decorrelation refers to the temporal difference between the observed post-fire backscatter coefficient change in regard to the time of change which may occur long after the vegetation combustion. In this chapter, different environmental variables that may influence radar scattering were analysed including, fire severity, post-fire vegetation recovery, soil, and vegetation water content, slope and topographic aspect. A random forests classifier was used to estimate the importance of these variables for the temporal decorrelation process. The analysis showed that over than 32% of the burned pixels located in the studied area were affected by temporal decorrelation, with fire severity, vegetation water content and soil moisture being its main drivers. However, when burned areas were detected long after fire, soil and the vegetation water content of both were the most important drivers behind the observed changes of the backscatter coefficient. The fifth chapter addresses the comparison between two burned area mapping algorithms based on radar images (Sentinel-1) and one based on optical (Sentinel-2). The analysis was carried out over ten study areas (10 million ha) in Africa. The algorithms were based on different mapping strategies and datasets (surface reflectivity, interferometric coherence and backscatter coefficient, the latter being presented in the second chapter). The maps were validated through reference perimeters derived independently from optical images (Landsat 8 and Sentinel-2). When considering all study areas, the optical data-based algorithm provided a significant increase in accuracy compared to radar-based ones. However, this may have been driven by the use of the same optical datasets when generating the reference fire perimeters (i.e., Sentinel-2). Nevertheless, the analysis suggested that optical image-based algorithms provide a significant increase in accuracy over radar-based algorithms. However, in regions with persistent cloud cover, the radar-based algorithms offered a valuable source of information, with radar-based detections being more accurate. In the sixth chapter, a comprehensive analysis of the use of convolutional neural networks (CNN) for burned areas detection and mapping is presented. CNN is a Deep Learning method widely applied in recent remote sensing studies. CNN were used to develop a seamless burned area mapping algorithm that includes radar (Sentinel-1), optical (Sentinel-2) and radar-optical datasets. Ten globally distributed study areas were considered. Five areas were used to establish the optimum dimensionality for feature extraction (i.e., 1D or 2D, data normalisation and the number of hidden layers). The remaining five areas were used to carry out an independent validation of the optimal models. Both the dimension and the optimal data normalisation were conditioned by land cover class and the senor type (optical or radar). The number of hidden layers only influenced the computation time without any improvements in the mapping accuracy being observed. The combination of radar and optical images allowed mapping burned areas with similar, or slightly higher accuracies when compared to those achieved in previous approaches developed within the Fire_cci program and based on both Sentinel-1 (DC 0.57) or Sentinel-2 (DC 0.7) datasets. Furthermore, the combined radar-optic approach eliminated information gaps due to the presence of clouds which affect detections based on passive sensors alone. In the seventh chapter, given the difference in accuracies between radar and optical data when mapping burned areas, various backscatter coefficient based temporal indices were evaluated to understand their suitability for burned areas mapping. The analysis was carried out to understand how radar-based burned area mapping algorithms may be improved in future work. The analysis is carried out using the random forests classifier, and the importance of each radar-based index is assessed when mapping burned areas. Depending on the land cover type, soil moisture, and topographic conditions, notable differences were observed between the temporal backscatter-based indices. In the eighth and last chapter, the conclusions derived from all research carried out within this doctoral thesis are resumed. The main findings, as well as the limitations found, are described as well as future lines of research that may help improving global mapping of burned areas from radar datasets and the combination of radar and optical datasets.