Remote Sensing for Mining Detection: Comparative Analysis of Deep Learning Models

Remote sensing combined with Deep Learning (DL) is gaining popularity for autonomous mining detection. However, it is observed that many DL models were assessed on custom datasets, which are not publicly accessible, with comparisons restricted to baseline models such as YOLOv5 and Faster R-CNN instead of other existing methods within the mining detection field. To address this gap, this paper provides a comparative analysis of six algorithms, including a baseline model (YOLOv5) and five algorithms derived from three recently proposed DL models consisting of cutting-edge algorithms such as convolutional neural networks, multiscale feature extraction, and attention modules. All six algorithms were executed and evaluated on a single, publicly available mining detection dataset, CUG_MISDataset. The results show that the MFTAYOLO algorithm demonstrated superior performance, achieving the highest mean average precision (mAP@50) with a 0.7 % improvement compared to other algorithms, alongside an approximate 17 % reduction in parameter count compared to the baseline model, YOLOv5.