Recognition of land use on open-pit coal mining area based on DeepLabv3+ and GF-2 high-resolution images

A highly efficient means is provided by remote sensing and deep learning to keep tracking of land use in open-pit coal mining area. Based on the high–resolution images from the domestic GF-2 satellite, a DeepLabv3+ model was utilized to achieve recognition of land use on open-pit coal mining area. In addition, a comparison was made among Deeplabv3+, U-Net, FCN, Random Forest, Support Vector Machine, and Maximum Likelihood Method. Firstly, samples data from high-resolution images were produced and sensitivity tests were conducted to determine the optimal cutting size and mode of the sample. Then, the deep neural network model (DeepLabv3+) was trained for conducting experiments of recognition of land use. Finally, the recognition results of different methods were compared. The results show that the optimal cutting size of the sample on the open-pit coal mining of the study area is 512 pixel×512 pixel. The optimal cutting mode of the sample is random cropping. The overall accuracy and Kappa coefficient of the DeepLabv3+ for recognition of land use on open-pit coal mining area are 80.10% and 0.73, respectively, which are better than the recognition accuracy of the U-Net, FCN, Random Forest, Support Vector Machine and Maximum Likelihood Method. The DeepLabv3+ is kept in the same order of magnitude as the above five methods. The feasibility of DeepLabv3+ model and GF-2 images in recognition of land use on open-pit coal mining area is verified, which is important for monitoring and restoration of eco-environment on open-pit coal mining area.