Change detection of open-pit mines based on FM-UNet++ and GF-2 satellite images

Automatic extraction of land use change information in open-pit mines using the remote sensing and deep learning technology is of great significance for the mining monitoring and ecological environmental protection. A novel deep learning model FM-UNet++ was constructed for the change of land use types in complex and heterogeneous mining scenarios, and the automatic change detection of open-pit mines was achieved using the Gaofen-2 (GF-2) satellite images. Firstly, the change detection dataset of open-pit mine was produced through data surveys and visual interpretation, which was augmented by data enhancement. Secondly, the FM-UNet++ for open-pit mine change detection was constructed by introducing the Feature-enhanced Coordinate Attention (FECA) mechanism and the Mish activation function into the UNet++. Finally, FM-UNet++ and 7 comparative models were trained for the change detection experiment of open-pit mine and the detection results of different deep learning models were compared. The results show that: (1) The FECA mechanism and the Mish activation function both improve the performance of the UNet++ for open-pit mine change detection. (2) The precision, recall, F1-Score and IoU of FM-UNet++ model are 95.6%, 89.2%, 92.3% and 85.7% respectively for the change detection of open-pit mine, which is significantly improved compared with the deep learning models such as FCN, PSPNet, Deeplab v3+, LANet, UNet, UNet++ and DA-UNet++. Besides, the change detection speed of the FM-UNet++ model remains in the same order of magnitude as the 7 comparative models mentioned above, with a stable training process, validating the feasibility of applying the FM-UNet++ model to the change detection in open-pit mines.