Problem Statement

The project addresses the challenges of detecting and monitoring unauthorized mining activities in resource-rich regions like Africa and Southeast Asia. These activities often cause severe environmental, social, and economic issues due to inadequate regulation and monitoring. The goal is to develop an accurate system for identifying mining sites using satellite imagery to support enforcement, conservation, and sustainable resource management.

Approach Used

The project employs Object-Based Image Analysis (OBIA) for identifying mining sites and integrates machine learning models to classify and segment mining areas. Sentinel-2 satellite imagery, with its multispectral capabilities, is analyzed to detect changes in the landscape indicative of mining activities. Key environmental indices and transfer learning techniques are used to enhance detection accuracy and adapt the model for varying geographic conditions.

Results

• Classification Model: Achieved an F1 score of 0.9902 on training and 0.8400 on validation, demonstrating high accuracy in detecting mining sites.
• Segmentation Model: Recorded Intersection over Union (IoU) scores of 0.5554 (validation) and 0.5458 (testing), highlighting robust segmentation capabilities.
• The project demonstrated the ability to accurately identify mining activities and provided actionable insights for conservation and regulatory enforcement.

Technology Used

• Models: MaxViT (a hybrid of CNNs and transformers) for classification and a binary segmentation model for precise delineation of mining areas.
• Techniques: Transfer learning from pre-trained models (e.g., ImageNet), OBIA for mine void identification, and multispectral analysis using Sentinel-2 satellite imagery.
• Performance Metrics: F1 score and IoU for evaluating classification and segmentation accuracy.
• Indices: Environmental indices such as spectral indices for detailed image analysis.