Fusion of Optical and Radar Data by Aggregation into a Single Feature Space for LULC Classification

Land use and land cover classification (LULC) is a fundamental input for ecological and socioeconomic models worldwide, generating a large volume of data from space-based platforms, mainly optical technologies. However, these can be affected by atmospheric conditions. Colombia has a high percentage of cloud cover due to its geographical location, which makes it challenging to map LULC changes. Studies have emerged that evaluate the integration of optical and radar images with algorithms that allow for good results despite the information gaps that affect these processes. Therefore, this work compares three supervised machine learning approaches, Support Vector Machines, Random Forest, and XGBoost, to classify land use and land cover from multispectral and radar images, contemplating four scenarios for data fusion. Optical, optical + SAR, optical + SAR ascending, and optical + SAR descending. The result for the Random Forest model using optical + ascending SAR has the best accuracy (76.02%), followed by Random Forest with optical + descending SAR data (75.97%) and with little difference for Random Forest using optical data (75.83%). In future studies, it is of great interest to explore feature extraction on both data sets to improve LULC representation and classification.