Supervised Classification of Water Bodies: Analyzing Various Methods Through a Case Study of Water Resources in Pakistan

In this paper, we explore the latest developments in the field of water body classification using satellite imagery, with a particular focus on Pakistani water resources as a case study. The spatial distribution of Earth's water and the change in its dispersion over time has critical implications for agriculture, ecology, and the environment. Classification of water bodies using remote sensing imagery is a significant research domain. Classical machine learning methods such as Random Forest, struggle due to their inability to capture intricate pixel-lovel associations within satellite imagery. However, deep learning and attention-based algorithms have emerged as strong contenders for effectively addressing this limitation. According to our information, no substantial work has been done in Pakistan at the national level aimed at employing Artificial Intelligence (AI) techniques for the classification and real-time monitoring of water resources. This study presents a comparative analysis of three techniques: Classical Machine Learning (ML) Algorithms, CNN-based Deep Learning Algorithms, and Attention-based Neural Networks. Leveraging a custom water dataset (PAKSEN2) with Sentinel-2 imagery, the study demonstrates that classical machine learning algorithms underperformed, achieving only 40% accuracy, while a CNN-based ResNet50 network [1] achieved 81%. Remarkably, the attention-based Vision Transformer (ViT) [2] achieved 99% accuracy, outperforming both traditional and CNN-based approaches.