CUG_MISDataset: A Remote Sensing Instance Segmentation Dataset for Improved Wide-Area High-Precision Mining Land Occupation Recognition

The effective and rapid acquisition of wide-area mine occupation information is crucial for ecological geo-environmental protection and sustainable development. Remote sensing instance segmentation technology based on deep learning is a promising solution. However, there are two significant challenges including insufficient training datasets and unsuitable segmentation models. To overcome these issues, this study provides a large-scale remote sensing instance segmentation dataset for mining land occupation (CUG_MISDataset). The CUG_MISDataset comprises 1426 image blocks and more than 3000 instances, covering all 150 types of mines found in China's Hubei province. It features multiple mine types, various land occupations, and complex instance scales. First, this study compares the performance of seven mainstream remote sensing instance segmentation models using the proposed CUG_MISDataset. The results show that all seven models achieve high segmentation accuracy. It indicates that the constructed CUG_MISDataset is robust and can serve as a valuable benchmark for remote sensing instance segmentation of mining areas. Second, aiming at the difficulty of large scale variation in this dataset, we propose a multiscale dilation feature pyramid network (MSD-FPN), which introduces a dynamic weight allocation mechanism to give more weight to important semantic information, while convolution with different dilation rates is used in the module to enhance the expression of mines’ multiscale features. The proposed MSD-FPN can achieve a 2.0% average precision improvement on the CUG_MISDataset.