Enhanced lithological mapping via remote sensing: Employing SVM, random trees, ANN, with MNF and PCA transformations

Abstract

This study examines the performance of three classification algorithms—Support Vector Machines (SVM), Random Trees (RT), and Artificial Neural Networks (ANN)—applied to Landsat 9 and Sentinel-2 spectral data for lithological mapping. The study area, located in the Central Anti-Atlas, is covered by the 1:50,000 geological map of Aït Semgane, featuring diverse geological formations, ideal for testing advanced remote sensing techniques. Results show that SVM, particularly with Minimum Noise Fraction (MNF) transformation, offers the best performance. For Sentinel-2 images, SVM with MNF achieves high user and producer accuracies and well-defined lithological boundaries. While RT and ANN also show good performance, they are slightly inferior to SVM, with RT achieving a Kappa index of 0.84 for raw Landsat 9 bands and ANN obtaining a maximum of 0.75 for Sentinel-2 data transformed with MNF. The MNF transformation generally improves SVM and ANN performance, whereas Principal Component Analysis (PCA) often produces inferior results. The robustness of SVM for high-dimensional data and its resistance to overfitting make it a promising tool for accurate lithological classification. This research has practical implications for geology and Earth sciences. The use of dimensionality reduction, particularly MNF, can greatly enhance classification quality for multispectral and hyperspectral data. These results are not only valuable for improving geological mapping, mineral exploration, and natural resource management at local and regional scales but also have significant potential for large-scale terrain analysis in diverse global contexts. The findings could support global efforts in geological hazard assessments, resource management, and environmental monitoring, particularly in regions with challenging geological settings. The study also proposes future research directions, such as exploring new dimensionality reduction techniques, evaluating classification methods with different remote sensing datasets, and integrating geophysical or geochemical data to further improve accuracy