Integration of spectral, thermal, and textural features of ASTER data using Random Forests classification for lithological mapping

Abstract

The ensemble classifier, Random Forests (RF), is assessed for mapping lithology using the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery over an area in southern Iran. The study area in the northern part of Rabor in the Kerman's Cenozoic magmatic arc (KCMA) is well exposed and contains some copper mineralization occurrences. In this research, the following six groups of ASTER datasets were used for RF classification: nine spectral bands in the VNIR and SWIR, five thermal bands in TIR, all 14 bands (including VNIR, SWIR, and TIR), band ratios, texture features, and principal components (PCs). The results showed that band ratios and all ASTER bands were able to more efficiently discriminate rock units than PC and texture images. The overall classification accuracies achieved were 62.58%, 55.40%, 65.04%, 67.12%, 54.54%, and 53.99% for the nine VNIR/SWIR bands, five TIR bands, all ASTER bands, band ratios, textural, and PCs datasets, respectively. Four datasets including all ASTER bands, band ratios, textural, and PCs datasets (37 bands) were combined as one group and applied in second RF classification which led to increase overall accuracy (up to 81.52%). Based on the four classified maps, an uncertainty map was produced to identify areas of variable (uncertain) classification results, which revealed that approximately 21.43% of all pixels on the classified map were highly uncertain. The RF algorithm found that 12 of the predictors were more important in the classification process. These predictors were used in a third RF classification, which resulted in an overall classification accuracy of 77.21%. Thus, the third RF classification decreases the accuracy. Field observations were used to validate our classification results.