Fine-Grained Urban Land Use and Land Cover Classification Through Multi-temporal and Multi-spectral Remote Sensing Images

Abstract

Earth observation via satellites using remote sensing techniques has been a long-standing research topic and it has observed popular attention recently due to the increase in accuracy rates with advances in deep learning techniques. In this paper, land use/cover map production has been conducted for the cities of Eskisehir and Kutahya and rural settlements located in Sakarya basin in Turkey. In detail, we used the CORINE 2018 data as ground truth, up to 11 urban classes, Random Forest as shallow classifier, and ResNet and DenseNet for CNN models and compared the results. In this study, using Red, Blue and NIR bands and 6 urban classes, 63% average accuracy and 0.53 kappa value have been observed as our top results.