Experimentations with applications of machine learning algorithms such as random forest (RF), support vector machines (SVM) and fuzzy inference system (FIS) to lithological classification of multispectral datasets are described. The input dataset such as LANDSAT-8 and Advanced Space-borne Thermal Emission and Reflection Radiometer (ASTER) in conjunction with Shuttle Radar Topography Mission (SRTM) digital elevation are used. The training data included image pixels with known lithoclasses as well as the laboratory spectra of field samples of the major lithoclasses. The study area is a part of Ajmer and Pali Districts, Western Rajasthan, India. The main lithoclasses exposed in the area are amphibolite, granite, calc-silicates, mica-schist, pegmatite and carbonates. In a parallel implementation, spectral parameters derived from the continuum-removed laboratory spectra of the field samples (e.g., band depth) were used in spectral matching algorithms to generate geological maps from the LANDSAT-8 and ASTER data. The classification results indicate that, as compared to the SVM, the RF algorithm provides higher accuracy for the minority class, while for the rest of the classes the two algorithms are comparable. The RF algorithm effectively deals with outliers and also ranks the input spectral bands based on their importance in classification. The FIS approach provides an efficient expert-driven system for lithological classification. It based on matching the image spectral features with the absorption features of the laboratory spectra of the field samples, and returns comparable results for some lithoclasses. The study also establishes spectral parameters of amphibolite, granite, calc-silicates, mica-schist, pegmatite and carbonates that can be used in generating geological maps from multispectral data using spectral matching algorithms.
Spectroscopic analysis is carried out for lithological discrimination in a study area in the Pali and Ajmer districts of Rajasthan, western India using laboratory-based, field-based and space-borne data. We first explored the feasibility of the Landsat-8 (Operational Land Imager (OLI), Thermal Infrared Sensor (TIRS)) and Advanced Space-borne Thermal Emission and Reflection Radiometer (ASTER) imagery for lithological mapping. Laboratory spectra of the samples of rocks exposed in the area were collected using FieldSpec3 spectroradiometer in the VNIR-SWIR region and resampled to the LANDSAT 8 and ASTER spectral bands. The spectral angle mapper (SAM) algorithm was used to map the lithologies using the resampled laboratory spectra as references. The resulting map was validated based on field geological mapping. Fourier Transform Infra-Red (FTIR) spectroscopy of selected rock samples was carried out to correlate spectral absorption features in the TIR-IR regions with the vibrational energy.