Advanced land imager superiority in lithological classification utilizing machine learning algorithms

Abstract

Different types of remote sensing data are commonly used as inputs for lithological classification schemes, yet determining the best data source for each specific application is still unresolved, but critical for the best interpretations. In addition, various classifiers (i.e., artificial neural network (ANN), maximum likelihood classification (MLC), and support vector machine (SVM)) have proven their variable efficiencies in lithological mapping, yet determining which technique is preeminent is still questionable. Consequently, this study aims to test the potency of Earth observing-1 Advanced Land Imager (ALI) data with the frequently utilized Sentinel 2 (S2), ASTER, and Landsat OLI (L8) data in lithological allocation using the widely accepted ANN, MLC, and SVM, for a case study in the Um Salatit area, in the Eastern Desert of Egypt. This area has a recent geological map that is used as a reference for selecting training and testing samples required for machine learning algorithms (MLAs). The results reveal (1) ALI superiority over the most commonly used S2, ASTER, and L8; (2) SVM is much better than MLC and ANN in executing lithologic allocation; (3) S2 is strongly recommended for separating higher numbers of classes compared to ASTER, L8, and ALI. Model overfitting may negatively impact S2 results in classifying small numbers of targets; (4) we can significantly enhance the classification accuracy, to transcend 90% by blending different sensor datasets. Our new approach can help significantly in further lithologic mapping in arid regions and thus be fruitful for mineral exploration programs.