Development of a Swin-Unet-Based Classifier Forthree-Class (Sediment Movement Area, Main Objects, Others) Recognition in Aerial Photographs

Key Points

• The aim is to improve the accuracy of recognizing sediment movement areas in aerial photographs by utilizing a Swin-Unet-based classifier.
• Developed a model for three-class classification (sediment area, main objects, others) using geographical information.
• Compared performances of classifiers with Swin-Unet, UNet, and UNet++ architectures.
• Implemented a slide partitioning method for input images and linear combination for output images to enhance segmentation accuracy.
• The Swin-Unet model demonstrated better accuracy in detecting sediment areas compared to existing approaches.
• The new linear combination method resulted in smoother and more precise detection boundaries.