Monitoring mining-induced subsidence from satellite imagery using transformer-based deep learning trained on gridded...

Area of Science:

• Environmental Science
• Geoscience
• Artificial Intelligence

Background:

• Underground coal mining poses challenges for environmental monitoring due to its hidden nature.
• Illicit mining activities often result in severe environmental damage, such as flooding and structural damage, before detection.
• Current methods lack early intervention capabilities, making ecological restoration difficult.

Purpose of the Study:

• To address limitations in AI model training data for detecting mining-induced land subsidence.
• To overcome the unsatisfactory classification accuracy of traditional Convolutional Neural Network (CNN) architectures in this domain.
• To develop a more effective AI model for monitoring land subsidence using satellite imagery.

Main Methods:

• A Vision Transformer (ViT-Base) model was trained using a large dataset of 191,630 land subsidence grid measurements and high-resolution satellite images.
• The model's performance was evaluated on ten representative mining-affected areas across China's top coal-producing provinces.
• Each evaluation area provided 250 independent subsidence grid measurements paired with satellite imagery.

Main Results:

• The ViT-Base model achieved 94% overall accuracy in identifying land subsidence within the training data region.
• Generalizability testing across diverse regions yielded accuracies ranging from 77.2% to 84.8%.
• The ViT-Base model demonstrated superior performance compared to conventional CNN models.

Conclusions:

• The Vision Transformer (ViT-Base) model offers a practical and accurate solution for monitoring mining-induced land subsidence.
• This AI approach overcomes data limitations and improves detection accuracy in various geological settings.
• The developed model provides a valuable tool for environmental authorities in managing and mitigating mining impacts.