Improving the Land Cover Mapping Accuracy of the Sentinel-2 Imagery on Google Earth Engine

Abstract

Accurate land cover information is necessary for successful monitoring, planning and management of the land cover features. Thanks to free-access satellite images, studies have focused on the creation of more accurate thematic maps, which have been used as a base data in many applications. The cloud-based Google Earth Engine (GEE) service makes it easier to access, store and process these satellite images. This study aims to improve the accuracy of a land cover map produced with the Sentinel-2 satellite image. For this purpose, as the very first step, study site was classified using only traditional bands of the Sentinel-2 data. To improve the classification accuracy, Sentinel-1 Synthetic Aperture Radar (SAR) data, Advanced Land Observing Satellite (ALOS) World 3D data, various spectral indices and gray-level co-occurrence matrix (GLCM) features were added to the traditional bands of the Sentinel-2 data, leading to a multi-source classification process. In this study, where the city center of Trabzon was selected as the study area, the accuracy of the land cover map produced using the Random Forest (RF) classification algorithm was increased from 83.51% to 92.78% with the auxiliary data used.