In this master’s thesis, we examine advanced satellite imagery methods for wildfire burned-area detection in Slovenia. Ten wildfire events serve as case studies to demonstrate the general applicability of the methods. Sentinel-2 satellite data were used, and a separability analysis of 192 spectral indices was performed to identify those with the highest ability to differentiate burned areas. The top-performing indices were averaged to create a new synthetic burned-area detection index, which served as a basis for several thresholding methods and a region-growing algorithm. In addition, learning samples and the best performing indices were used for machine learning with a random forest classifier. All methods were applied to both pixel-based and object-based image analysis, the latter requiring prior image segmentation. The methods were evaluated using confusion matrices, based on a comparison of the detected burned areas with official burned-area polygons provided by the Slovenia forest service. Our results indicate that the pixel-based region growing algorithm achieved the highest burned-area detection accuracy.
|
