Abstract
Monitoring changes in lake area using remote sensing imagery and artificial intelligence algorithms is essential for assessing regional ecological balance. However, most current semantic segmentation models primarily rely on the visible light spectrum for feature extraction, which fails to fully utilize the multi-spectral characteristics of remote sensing images. Therefore, this leads to issues such as blurred segmentation of lake boundaries in the imagery, the loss of small water body targets, and incorrect classification of water bodies. Additionally, the practical applicability of existing algorithms is limited, and their performance under real-world conditions requires further investigation. To address these challenges, this paper introduces SCR-Net, a water body identification model designed for multi-spectral remote sensing images. SCR-Net employs a dual-channel encoding-decoding mechanism and alters the number of channels used for reading image data, enhancing feature learning for lakes while focusing on extracting information about the water body target locations, thereby ensuring accurate segmentation. Trained on multi-spectral remote sensing images, the model leverages the unique spectral properties of these images to improve segmentation accuracy. Extensive validation on two datasets demonstrates that SCR-Net outperforms state-of-the-art models in terms of segmentation accuracy. Based on the validation using this dataset, Daihai Lake in Inner Mongolia was additionally selected as a case study to calculate the lake area, providing valuable insights for interdisciplinary research in ecological environment monitoring and remote sensing image processing.