Abstract: The monitoring of biological soil crust can help the effective development of surface ecological restoration in mining areas, and obtaining biological soil crust information based on UAV remote sensing has attracted wide attention. It is difficult to extract biological soil crust from images because of the scattered and uneven growth of biological soil crust on the surface of mining area and the complexity of field environment. Based on the high precision and high efficiency of deep learning model classification, an improved UNet++ model and a lightweight DeeoLabV3+ model were proposed to extract biological soil crust. Firstly, the best Epoch, Backbone and loss function are optimized based on the network structure of UNet++ model, and the improved UNet++ model with ResNeXt as the Backbone and Dice Loss as the loss function is obtained. Compared with the test results of UNet++ and U-Net, the results show that the improved UNet++ model has the best classification effect. The accuracy rate of biological soil crust is 97.88%. Then, based on DeepLabv3+ model, the original backbone network Xception network of DeepLabv3+ is changed into MobileNetV2 lightweight network, and the expansion rate of ASPP module in the original DeepLabv3+ were optimized and improved. Compared with the original DeepLabv3+ model and the model that transforms DCNN structure of DeepLabv3+ into MobileNetV2 network, the results show that the training time of the proposed lightweight DeepLabv3+ model is one quarter of that of the improved UNet++ model, and the size of the model is nearly one tenth of that of the improved UNet++ model, which can extract biological soil crust quickly and accurately. In terms of accuracy, the improved UNet++ has obvious advantages in extraction accuracy. Lightweight DeepLabV3+ model has greater advantages in training time and model size. The proposed method is suitable for extracting biological soil crust with uneven growth and complex environment, and has achieved good results, which provides good data support for studying the development law of biological soil crust in arid areas, and further explores the application of deep learning model in the research field of surface information acquisition in mining areas.