WaterFormer-MS: A multi-stream deep learning framework for surface water segmentation from satellite imagery

Abstract

Accurate mapping of surface water bodies is essential for dam monitoring, irrigation planning, and sustainable water resource management. Conventional methods based on spectral indices, such as NDWI, can face challenges in complex environments due to spectral ambiguity, seasonal variability, and urban or vegetated surfaces. Deep learning approaches such as U-Net and DeepLabv3+ have improved water-body segmentation; however, accurately delineating fine boundaries and detecting small water bodies remain challenging. To address these limitations, this study proposes WaterFormer-MS, a multi-stream deep learning framework for surface water segmentation from Sentinel-2 imagery. The architecture integrates spectral, geometric, and contextual information through dedicated processing streams and combines the extracted features using a convolutional fusion module enhanced with CBAM attention. This design aims to improve feature representation and support accurate water-body delineation. The proposed model was trained and evaluated using Sentinel-2 imagery from a publicly available Kaggle dataset containing diverse land-cover conditions, including urban, vegetated, and agricultural environments. At 300 epochs, WaterFormer-MS achieved an overall accuracy of 97.4%, a precision of 93.2%, a recall of 94.2%, an F1-score of 93.7%, and an IoU of 89.7%. Within the experimental framework, WaterFormer-MS achieved higher values than the other proposed variants. Comparisons with conventional NDWI-based approaches and published deep learning methods suggest competitive performance; however, these comparisons should be interpreted cautiously because datasets, preprocessing procedures, training protocols, and evaluation settings may differ. The results support the potential of the proposed framework for surface water segmentation and environmental monitoring applications in diverse remote sensing scenarios.