Deep Learning based Minerals’ Recognition and Mapping using Sentinel-2 Imagery

Abstract

Abstract Marble and limestone possess calcium carbonate (chemical formula: CaCo3) as major ingredient that’s why they are called carbonates or carbonated mineral. Carbonates are 70% of the total minerals’ deposits of the intended study area which are divulged and mapped using the significant deep learning neural network models and latest Sentinel-2 imagery. While delineating them, an overall accuracy of 96% for 1-dimensional convolution neural network and 95% for artificial neural network was achieved while targeted carbonates class accuracy remained 99% and 100% respectively. Sentinel-2 sensors record data in visible, Near infrared, and short wave infrared bands which are much appropriate to delineating carbonated minerals as they show greater absorption features in these bands. Sentinel-2 data was downloaded in Level-2 format and resampled to 10 meter spatial resolution using bilinear nearest neighbors algorithm. Significant amount of data polygons (2500+) were drawn and cleaned up for various class members in order to prepare them for various deep learning and machine learning models. Data was split in the ratio of 70:30 as training-test sets which provided with the most optimal mapping results. Classification and accuracy assessment reports of the models with high quality resultant imageries were overlaid in ArcGIS 10.2 and presented in article.