A Comparative Study of Machine Learning Models for Pixel-Wise Acid Mine Drainage Classification Using Sentinel-2

Acid Mine Drainage (AMD) is a type of water pollution that occurs when sulfur-bearing minerals in rocks and soil are exposed to air and water. This paper explores the application of Machine Learning (ML) algorithms for pixel-level classification of AMD from Sentinel-2 imagery. The proposed algorithms include Random Forest (RF), K-Nearest Neighbor (KNN), Logistic Regression (LR), and Multi-Layer Perceptron (MLP). They classify each pixel in the imagery as AMD or non-AMD. To generate more training samples and improve the performance of ML methods, we use a window-based approach and the SMOTE technique [1]. Best hyperparameter of ML methods are selected based on the greedy forward selection approach. The study also investigates the relative importance of different spectral bands of Sentinel2 for AMD classification. The experimental results are collected on three lakes in Outokumpu mining region, Finland. The results demonstrate that RF model exhibits the best classification performance with accuracy of 97.74%.