[Objective] Based on deep learning, a large-scale and high-precision positive and negative terrains segmentation is realized, and the effective segmentation of positive and negative terrains has important theoretical value and guiding significance for soil erosion control and ecological restoration and reconstruction in the Loess Plateau. [Methods] The typical sample area was selected in the hilly area of the Loess Plateau, and the terrain segmentation data set was made by using the medium resolution DEM data. The positive and negative terrains segmentation model of the improved Unet was constructed. Based on the Unet model structure, the residual module was introduced to replace the convolution module to deepen the network structure and increase the extraction of terrain information. Combined with the convolution attention module, eliminating useless information increased the anti-interference of the model; the activation function and loss function were optimized to enhance the robustness and accuracy of the model. [Results] The overall accuracy of the terrain segmentation of the slope deformation neighborhood judgment method is 70.3%. Among the deep learning models, the improved Unet model has the best effect, with a certain improvement compared with both the Unet model and the Res-Unet model, and the overall accuracy reaches 86.2%. [Conclusion] Compared with the traditional slope distortion neighborhood judgment method, the accuracy evaluation index of the network model segmentation results based on deep learning is better, which verifies the effectiveness of the improved Unet model in terrain segmentation. [ABSTRACT FROM AUTHOR] more...