Interactive Segmentation Method Utilizing Spectral and Contour Features: A Case Study on a Dual-Task Remote Sensing Water Body Dataset

This study introduces a semi-automatic approach designed to enhance the process of labeling water body (WB) by effectively leveraging both spectral and contour characteristics. By incorporating innovative techniques, such as anchor-livewire and region of interest (ROI)-livewire algorithms, we address the challenges associated with incomplete or missing labels encountered in traditional approaches, such as remote sensing indexes, or interactive segmentation method, such as LabelMe. This enables rapid and precise labeling of WB in remote sensing images. To demonstrate the efficacy of our approach, we conducted not only simulated experiments but also applied it to label a comprehensive dual-task WB dataset named Water-CD. This dataset consists of 4596 pairs of Sentinel-2 dual-phase remote sensing images accompanied by highly accurate WB labels and WB change labels. Additionally, our contribution significantly addresses the gap in publicly available datasets for WB change detection. We evaluated the labeled dataset in both WB segmentation and WB change detection tasks using various deep learning methods, conducting extensive experiments to assess performance. The results indicate that deep learning models trained on labels generated by our method outperform those trained on automatically generated labels using the normalized difference water index (NDWI) in the same test set. Specifically, we observed a significant 28.84% improvement in the F1 score in the WB change detection task, which focuses more on precise contours. This underscores the effectiveness and superiority of our approach in accurately labeling WB. The Water-CD dataset can be accessed and downloaded from
https://github.com/ChogoriPeak/Water-CD
.