Classification of Water Bodies Using Ensemble of U-Net and Random Forest Algorithm

Abstract

Water bodies classification using remote sensing and deep learning techniques plays a pivotal role in the effective management of water resources. This study aims to address the challenge of accurate water body detection and classification, which is essential for understanding their distribution and characteristics, ultimately informing water usage and conservation efforts. Current methodologies predominantly rely on Support Vector Machines (SVM) and pixel-based approaches, resulting in suboptimal accuracy. In response, this paper proposes an ensemble model that combines the U-Net neural network and the Random Forest algorithm for enhanced water body detection and classification. The research commences by obtaining high-resolution satellite images with a resolution of 0.5 m. The U-Net model is employed to segment water bodies, and contour analysis is subsequently applied to extract shape features. The Random Forest Classifier is then utilized to classify the segmented water bodies into distinct categories, including rivers, ponds, lakes, canals, and other water bodies. Following the U-Net segmentation, the rasterized segments are converted into vector format. These vector data are leveraged to update Geographic Information System (GIS) maps, contributing to more accurate cartographic representations. The proposed approach is rigorously evaluated using a dataset from urban areas in Kolkata, West Bengal, India. The achieved accuracy rate stands at 67.01%.