Extraction and analysis of riverbank lines in Khulna district of Bangladesh based on an improved TransUNet model

This study proposes a novel riverbank line extraction method by integrating a Dynamic Channel Graph Convolutional Network (DCGCN) with TransUNet. The DCGCN module captures local semantic structures and adaptively enhances inter-channel feature interactions, improving the detection of small tributaries and complex river systems. A riverbank line dataset from Khulna District, Bangladesh, was constructed to evaluate the proposed model. Experimental results demonstrate excellent performance, achieving an overall accuracy of 99.36%, IoU of 90.62%, MIoU of 94.97%, FWIoU of 98.75%, and Dice coefficient of 95.08%. Riverbank lines from 2017 to 2024 were further extracted and analyzed using the Digital Shoreline Analysis System (DSAS) to assess temporal variation and migration trends. The results indicate an overall slight progressive trend, with pronounced local erosion at major river bends. The proposed framework provides an effective tool for accurate riverbank mapping and quantitative river system monitoring.