MACHINE LEARNING-BASED SNOW COVER MAPPING IN UTTARKASHI, CHAMOLI, AND PITHORAGARH USING CLOUD BASED REMOTE SENSING TOOL

Abstract

Snow cover monitoring is essential for hydrological modelling, climate change analysis, and water resource management, especially in the Himalayan cryosphere. The most cutting-edge global open-source platform for sophisticated geospatial big data analysis is Google Earth Engine (GEE). This study leverages Google Earth Engine (GEE) and data sets available, that is Harmonized Sentinel-2 imagery, VIIRS, and Digital Elevation to delineate annual snow cover in Uttarkashi, Chamoli, and Pithoragarh districts of Uttarakhand.  This paper aims to (i) Land Use Land Cover (LULC) Mapping. (ii) Detection of Snow cover in the Himalayan region districts of Uttarkashi, Chamoli, and Pithoragarh, Uttarakhand, India, using the annual composite median of Sentinel-2 imagery. (iii) To compare the performance of various machine learning models, that is Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Tree (CART) for 5 classes. (iv) To calculate the area of 5 classes for the years 2019 and 2024. (v) To build classified maps using the algorithm that results in the best overall accuracy.  Here, three machine Learning approaches Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Tree (CART) are trained using input parameters such as bands, spectral indices (NDVI, NDBI, NDSI, BSI), and topographic parameters (elevation, slope) derived from ALOS DEM. Cloud-masking techniques refine the dataset, ensuring high-quality spectral inputs. The result demonstrated the successful mapping of LULC's five land cover classes: bare soil, snow, vegetation, built-up areas, and water bodies. The study demonstrated high classification accuracy in 2019 for RF, SVM, and CART across all districts, achieving 95.7%, 93.2%, and 90.7% in Chamoli; 96.5%, 97.3%, and 95.6% in Pithoragarh; and 88.6%, 90.0%, and 87.3% in Uttarkashi. In 2024, the accuracy rates improved to 96.2%, 93.9%, and 94.6% for Chamoli; 95.8%, 92.5%, and 91.6% for Pithoragarh; and showed significant gains reaching 95.4%, 95.4%, and 96.1% for Uttarkashi. Results indicated that estimated In Chamoli, RF consistently performed better, demonstrating an 8.3% increase in snow from 2,206 km² to 2,388 km², while Pithoragarh experienced a 25% loss from SVM to RF:  from 2,099 km² to 1,573 km²).Snowfall in Uttarkashi increased by 10.8% from SVM to CART: from 1,804 km² to 1,998 km2, with CART doing exceptionally well in 2024.
RF proved most reliable overall, but regional variability suggests need for adaptive model selection.