An Enhanced UNet Model to Detect Water Bodies From Remote Sensing Data

Water body segmentation is an important task in remote sensing analysis and image interpretation due to its use in various studies, such as land cover classification and disaster management. Developing an automated water segmentation approach is crucial for a better extraction of water bodies with minimal human intervention. This study demonstrates a modified version of UNet, a commonly used Convolutional Neural Network (CNN) for segmentation, by introducing attention mechanism into its architecture. Attention is achieved using Squeeze-and-Excitation (SE) mechanism, which is expected to enhance feature extraction and construction of segmentation maps. A dataset known as “Semantic Segmentation of Aerial Imagery” is used for training and testing purposes to verify the efficiency of the proposed network, called SE-UNet. The results are evaluated using Accuracy, Recall, Precision, F-score and Intersection over Union (IoU) metrics. Comparative evaluation and experimental results show that using SE to embed Attention mechanism into UNet significantly improves the overall performance.