Fine mapping of Hubei open pit mines via a multi-branch global–local-feature-based ConvFormer and a high-resolution benchmark

Abstract

As an important part of the geological environment assessment in recent years, mine interpretation has attracted broad attention in the area of sustainable development. As an object that can be captured by remote sensing of multispectral imageries, the open-pit (OP) mine environment has the characteristics of irregular distribution, diverse shapes, and varied textures and colors, which makes interpretation a challenge and limits the accuracy of OP mine mapping. Moreover, the development of computational vision technology enables efficient and fast interpretation. However, there is a lack of open semantic segmentation datasets for OP mining areas. In this paper, we introduce a dataset for OP mining areas in the West Hubei metallogenic belt (DOWHM) for OP mine element interpretation. DOWHM includes 24 active OP mining zones in China’s Hubei Province, with a total of 1875 satellite images of 224×224 pixels at a resolution of 0.6 m. Second, we design a ConvFormer model for fine classification and mapping in OP mining areas. The ConvFormer model adopts an encoder featuring a large convolutional kernel. The decoder of the ConvFormer model employs a multi-branch approach, integrating modules of convolutional neural network (CNN) and self-attention mechanism. In this manner, ConvFormer effectively captures both global–local information. This solves the problem of the misclassification of elements in the prediction results, and especially has advantages in road classification. In the experiments, ConvFormer achieved mIoU of 68.18% on DOWHM. Compared to other popular semantic segmentation models, ConvFormer improved the overall mIoU by 2%, which is advantageous in the fine classification of OP mining regions, and it can be applied to the mapping of entire mining regions. DOWHM and the code are available at https://github.com/zxy1211/DOWHM.

Keywords

• Open-pit mining area
• Remote sensing
• High resolution remote sensing image
• Semantic segmentation
• Self-attention mechanism