Machine Learning-Based Improved Land Cover Classification Using Google Earth Engine: Case of Atakum, Samsun

Abstract

Land cover (LC) mapping using remote sensing images is essential in studies such as environmental management, urban planning, ecological research, etc. The study aims to produce a classified land cover map of the Atakum district using machine learning methods in a Google Earth Engine (GEE) environment. Random Forest (RF) and Gradient Tree Boosting (GTB) methods were used in the study. Landsat 8 satellite images and ALOS DEM were used as datasets. Normalized Difference Vegetation Index (NDVI), Normalised Difference Building Index (NDBI), Normalised Difference Water Index (NDWI), Bare Soil Index (BSI), Soil-Adjusted Vegetation Index (SAVI), and Enhanced Vegetation Index (EVI) were used to improve the classification. The land cover in the study area was classified as impervious, vegetation, farmland, barren land, and water bodies. All input variables were normalized to optimize the performance of the model. The performance of the model was evaluated using user accuracy, producer accuracy, overall accuracy, and kappa coefficient accuracy evaluation techniques. In this study, the calculated kappa coefficients of RO and GTB for the prepared land cover are 95.6% and 96.0%, and the average overall accuracy is 96.8% and 97.1%, respectively. In the study, it was observed that GTB outperformed RO among the two machine learning methods.