Deep and shallow feature fusion framework for remote sensing open pit coal mine scene recognition

Area of Science:

• Environmental Science
• Remote Sensing
• Computer Vision

Background:

• Effective monitoring of open-pit coal mining areas is vital for environmental management.
• Existing land use and damage recognition methods face limitations, including manual feature design and heavy reliance on sample data.

Purpose of the Study:

• To develop an advanced scene recognition framework for remote sensing images of mining areas.
• To overcome limitations of traditional and deep learning methods by fusing deep and shallow features.

Main Methods:

• A three-branch feature extraction framework combining deep features (DF) and shallow features (SF).
• Deep features enhanced using a neighboring feature attention module and Graph Convolutional Network (GCN).
• Shallow features extracted via Gray-Level Co-occurrence Matrix (GLCM) and Gabor filters, fused and classified using particle swarm algorithm optimized support vector machine (PSO-SVM).

Main Results:

• The deep feature extraction model achieved high accuracies of 97.53% (AID) and 96.73% (RSSCN7).
• The fused feature classification using PSO-SVM reached an accuracy of 92.78%.
• The proposed method outperformed four other classification approaches.

Conclusions:

• The proposed framework effectively fuses deep and shallow features for accurate remote sensing scene recognition.
• This approach enhances land use and damage assessment in mining areas with fewer samples.
• The study demonstrates a superior method for environmental oversight and management in mining regions.