Surface Water Mapping Using Machine Learning and Deep Learning Techniques from LISS-III Imagery

Surface water mapping is crucial for preserving and strategically managing water resources. However, extracting water bodies and estimating their extent from satellite images faces challenges due to the similar spectral responses of different land types. The task of identifying optimal features from satellite images and designing deep neural networks with reduced computational complexity is also demanding. In this paper, the Machine Learning (ML) with conventional, Pixel Level Feature (PLF), and Multi Scale Feature Extraction Network (MSFEN) feature extraction methods and Lightweight Attention-based Multiscale Neural Network (LWAMNNet) are compared for surface water mapping using satellite images. Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGB) are the ML classifiers considered for this study. The Linear Imaging Self-Scanning Sensor-III (LISS-III) images from the Resourcesat-2 satellite are used for experimentation. The results indicate that LWAMN-Net outperforms traditional ML techniques. Among the ML techniques, the MSFEN feature extraction method demonstrates superior performance, particularly when combined with SVM (MSFEN-SVM), outperforming other ML techniques.