Open-pit Mining Area Segmentation of Remote Sensing Images Based on DUSegNet

Remote sensing is an important technical means for monitoring and protecting mineral resources. However, because of the complex surface environment, very few good results have been achieved in the study of automatic open-pit mining area segmentation. Inspired by SegNet, UNet and D-LinkNet, this paper proposes a novel deep convolutional neural network for pixel-level semantic segmentation of optical remote sensing images termed DUSegNet. In this network, the pyramid model and upsampling method of pooling indices, similar to SegNet, are employed in the encoder-decoder architecture. In addition, the convolutional skip connection architecture, similar to UNet, is adopted to connect shallow features to the decoder. Additionally, the serial-parallel model, similar to D-LinkNet; the intensifier constructed by dilated convolution; and the classifier constructed by softmax layers are applied in the process of encoding and decoding. In the practical application stage, we present an effective open-pit mining area segmentation method for entire remote sensing images, which has great significance for practical work, such as environmental impact assessment procedures and mine management. In the experimental stage, we compared the open-pit mining area segmentation effects of SegNet, UNet, DecovNet, and DUSegNet on the same dataset manually collected from GF-2 remote sensing images and verified the advantages of DUSegNet using graphic results and optimal evaluation metrics, such as AP (0.94) and F-score (0.67).

Automated summary

This paper proposes a novel deep convolutional neural network DUSegNet for pixel-level semantic segmentation of optical remote sensing images. Experimental results show that DUSegNet outperforms other algorithms in open-pit mining area segmentation, demonstrating practical significance.