High-resolution mapping of seasonal snow cover extent in the Pamir Hindu Kush using machine learning-based integration of multi-sensor data

This study proposes a framework to develop a high-resolution snow cover area (SCA) product from freely available spaceborne remote sensing data and utilizes the Sentinel-1 multi-temporal products and MODIS surface reflectance data. The proposed methodology focuses on using the sensitivity of the parameters retrievable from the Sentinel-1 datasets to snow. Different parameters such as the dual polarimetric entropy, mean scattering angle, backscatter coefficients, and the interferometric coherence are integrated with a spatially resampled normalized difference snow index (NDSI) from MODIS data to estimate an equivalent NDSI, which is used for the determination of the SCA at 15 m spatial resolution. The equivalent NDSI is derived using a machine learning-based regression based on support vector machines (SVMs) and the multilayer perceptron (MLP). The experiments are performed for the high elevated regions of the Kunduz and Khanabad watershed of the northern Hindu Kush mountains for the peak winter and early melt season of 2019, corresponding to February and March. The reference SCA for evaluating the results is generated by thresholding the NDSI derived from pan-sharpened Landsat-8 imagery. As compared to MLP, the SCA generated based on the SVM regression showed better performance. Further, compared to spatially resampled MODIS NDSI, both the SVM and MLP results showed better accuracy for snow classification, as determined by the mean conditional kappa coefficients of 0.75, 0.83, respectively, over 0.62.