Monitoring of Spatio-Temporal Snow Cover using AI Based Self-Supervised Learning in Data Scarce Himalayan River Catchment

Seasonal snow cover in high mountain regions plays an important role in river flow hydrographs, water availability and land atmosphere interactions. In the Western Himalaya, mapping snow cover accurately is difficult due to frequent cloud cover, steep terrain, strong shadow effects and confusion between snow, clouds and bright rocky surfaces in satellite images. The study shows that these issues are clearly observed in the Western Himalayas, where the lack of sufficient ground reference data further limits the use of fully supervised classification methods. To address these challenges, the objective of the present study is to develop a snow cover mapping framework that leverages self-supervised learning for robust feature representation from satellite remote sensing data in the Western Himalaya. The method focuses on learning useful snow related features directly from large volumes of unlabelled satellite images, reducing the need for extensive manually labelled training data. Multi temporal optical satellite images are used so that the model can learn stable snow patterns across different seasons, illumination conditions and surface states. A convolutional neural network is trained using a contrastive self-supervised learning strategy, where different augmented versions of the same image patch are treated as similar samples, while patches from different locations are treated as dissimilar. The self-supervised encoder is coupled with a lightweight decoder in an encoder-decoder segmentation architecture, enabling pixel wise snow mapping while preserving spatial detail under limited supervision. Simple data augmentations, such as brightness changes, contrast adjustments and random cropping are applied to improve the model’s ability to recognize snow under varying conditions while preserving its key spectral and spatial characteristics. After self-supervised pretraining, the learned feature representations are fine tuned for snow and non-snow classification using a limited set of labelled samples derived from reference snow products and manual interpretation. This greatly reduces the dependence on large labelled datasets compared to conventional supervised learning methods. Snow cover maps are generated for different seasons and elevation zones to examine spatial and temporal variability of snow distribution across the basin. The results are compared with traditional index based methods, such as Normalized Difference Snow Index (NDSI) thresholding, especially in areas affected by clouds, shadows and mixed land cover. The study shows that the self-supervised learning provides a practical and reliable framework for snow cover mapping in data scarce and high altitude regions. The methodological framework developed in this study can be utilized for other basins also to have improved understanding of snow cover dynamics.

Keywords: Snow cover; self-supervised learning; remote sensing; Himalaya