Image recognition system for ecological restoration of high and steep slopes in open pit mines based on improved VGG16

• Published in: Discover applied sciences Vol. 7; no. 9; pp. 934 - 17
• Main Authors: Gong, Yongfeng, Zhou, Jianwei, Zhang, Gang, Li, Ran, Wang, Guorui, He, Xiaofeng, Hu, Zhiyong
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.09.2025; Springer Nature B.V; Springer
• Subjects: Accuracy; Algorithms; Applied and Technical Physics; Automation; CapsNet; Chemistry/Food Science; Deep learning; Earth Sciences; Ecological monitoring; Ecological restoration; Engineering; Environment; Environmental management; Environmental restoration; Image processing; Image recognition; Image segmentation; Materials Science; Mines; Mining engineering; Monitoring; Monitoring methods; Occlusion; Open pit mining; Remote sensing; Robustness; Seasonal variations; Semantic segmentation; Sustainable development; U-Net; Vegetation; VGG16; Image recognition; Ecological restoration; CapsNet; VGG16; U-Net; Mines
• ISSN: 3004-9261; 2523-3963; 3004-9261; 2523-3971
• DOI: 10.1007/s42452-025-07634-6

Ecological restoration of high and steep slopes (HSSs) in open pit mines (OPMs) is a key aspect of mining environmental management. However, relevant mine restoration image datasets are usually limited in size. Moreover, such images are susceptible to factors such as ambient light and occlusion, leading to inefficiency and subjectivity in traditional monitoring methods. To improve the accuracy and automation of slope restoration monitoring, an image recognition system combining improved visual geometry group 16-layer network (IVGG16), U-shaped network (U-Net), and capsule networks (CapsNet) is designed in this study. Meanwhile, in the baseline model U-Net, IVGG16 and CapsNet are introduced to build a semantic segmentation model to improve image recognition accuracy. The experimental results indicated that the mean intersection over union (mIoU) of only introducing CapsNet was 87.45%, which was improved by 2.50% compared with the baseline U-Net model. The mIoU of introducing only IVGG16 was then improved by 4.79–89.74%. Improving the VGG16 model enhanced feature extraction capability and effectively reduced overfitting. Whereas, CapsNet could capture the spatial hierarchical relationships, enhance the detail sensitivity, and optimize the spatial relationship modeling. The present model demonstrated excellent robustness under challenging conditions such as complex lighting and seasonal changes. For example, under rainy conditions, the fluctuation range of its mIoU could be controlled within 6.45%, and the model maintained a stable output, with a significantly better performance than traditional methods. In the complex scenario of mine slope rehabilitation, the model mIoU proposed in this study was 90.16% and F1 Score was 94.86%. It had good segmentation accuracy and robustness, and the performance was optimized. This provides reliable technical support for intelligent monitoring of ecological restoration and promotes the green and sustainable development of mining engineering.