Benggang Extraction Based on Improved U-Net Model from Satellite Remote Sensing Images

The Benggang is a typical and severe soil erosion type in the red soil region of southern China. Accurate identification and extraction of Benggang can provide a reliable basis for planning the treatment of the gullies. In this paper, an improved U-Net model is proposed, which uses a three-dimensional convolutional structure, introduces dilated convolution to replace the original max-pooling layer, and solves the problem of decreased image resolution caused by downsampling in image segmentation. The residual module is introduced to improve the convergence performance of the network and prevent gradient disappearance during the training process. In addition, an attention mechanism is adopted to enhance the expression of features in image segmentation by combining the spectral and spatial features of remote sensing images. Furthermore, some optimizations are made in the preprocessing and post-processing of the model to facilitate the model training and prevent overfitting. The study area of this paper is a concentrated area of Benggang in Jiangxi Province, China, and GF-2 satellite image data is used to compare the improved U-Net model with other mainstream models. The results show that the improved U-Net model can effectively extract the spectral and spatial features of images, facilitate the detection of the target to be segmented, and effectively prevent the overfitting and degradation problems of the network, enhancing the robustness of the network. The proposed method can accurately segment and extract Benggang from remote sensing images with high precision and strong robustness, providing reliable monitoring data for the prevention and treatment of Benggang.