Abstract
Abstract. Accurate estimation of groundwater levels in river basins is paramount for hydro-geological research and sustainable water resource management. In this paper, we introduce a deep learning framework explicitly developed for precise groundwater level estimation in the Ganga River Basin. Leveraging the combined band information of Sentinel-1 synthetic aperture radar (SAR) and GRACE satellite data, our approach capitalizes on the trans-formative capabilities of Vision Transformers (ViT) and their variants, with a particular focus on Swin-Transformer variant enriched with Normalization Attention Modules (NAMs).To address the unique challenges of the Ganga River Basin, we curated a comprehensive dataset, forming a robust foundation for training computer vision models tailored to this distinct geographical region. Through rigorous experiments, our state-of-the-art Vision Transformers demonstrated significant potential in groundwater level estimation, with the Swin-Transformer NAM-based model achieving an outstanding Mean Absolute Error (MAE) of 1.2. These remarkable results surpass conventional methodologies and underscore the substantial advancements achieved through advanced transformer-based architectures in this domain. Moreover, this research contributes a robust dataset for future endeavours, fostering further advancements in groundwater estimation and related fields. This study represents a substantial step towards advancing sustainable groundwater utilization practices in the Ganga River Basin and beyond.