Mapping snow avalanche debris by object-based classification in mountainous regions from Sentinel-1 images and causative indices

With the rapid development of satellite observation datasets, avalanche detection algorithms are not as accurate as visual interpretation, limiting avalanche hazard management. To bridge this gap, more advanced machine learning is proposed to map snow avalanche debris. Those techniques use Sentinel-1 SAR scattering characteristics and field observations with principal component analysis (PCA), support vector machine (SVM), and logistic regression (LR) in the western range of the Tianshan Mountains of Xinjiang, China. Specifically, the indicators in the snow avalanche debris samples described the time-shift variations, quantified by the variations from the ascending and descending image pairs. Then, combined with the causative factors, PCA-LR and PCA-SVM transformed point-monitoring at the regional scale. Finally, the snow avalanche debris distribution was detected (13.92 m). It was found that: (1) The accuracy of snow avalanche debris detection was not enhanced by ascending or descending image pairs. Although the ascending image results outweigh the descending ones, it underestimated the amount of debris with high miss and false detection rates. (2) The composite results of the ascending and descending adjacent image pairs were highly satisfactory for snow avalanche debris detection. Although the PCA-LR results narrowly overtook those for PCA-SVM (CSILR1 = 86.38 vs. CSISVM1 = 83.06, PODLR1 = 98.90 vs. PODSVM1 = 95.37; CSILR2 = 84.90 vs. CSISVM2 = 81.53, and PODLR2 = 98.56 vs. PODSVM2 = 94.15), both results overestimated the debris amounts (FBLR1 = 113.39 vs. FBSVM1 = 110.19; and FBLR2 = 114.64 vs. FBSVM2 = 109.64), with low miss and false detection rates (FARLR1 = 12.73 vs. FARSVM1 = 13.44; FARLR2 = 14.03 vs. FARSVM1 = 14.13). (3) False and missed detection of avalanche debris pixels occurred due to the SAR images' limitations and an incorrect signal from the massive, deep frost caused by thick snow. The high-accuracy approach using multiple orbits, polarizations, and terrain indices was encouraging because they revealed slab-and groove-type avalanche debris from noise filtering and speckle reduction.