A Review Based on Satellite-Based Land Cover Classification System

Abstract

Accurate land-cover information is essential for understanding environmental change, supporting sustainable development, and assisting planning in rapidly evolving urban and agricultural landscapes. Although satellite imagery is widely accessible, transforming raw multispectral data into reliable large-scale land-cover maps remains a non-trivial task due to data complexity and the expertise typically required for interpretation. This study presents an automated land use and land cover (LULC) classification framework that generates pixel-level thematic maps directly from Sentinel-2 Level-2A GeoTIFF imagery. The proposed system integrates selected spectral bands with reference annotations derived from the ESA WorldCover dataset and enhances the feature representation through domain-driven spectral indices such as NDVI, NDBI, and MNDWI. Training samples are obtained from geographically diverse regions and balanced across categories to improve robustness and generalization. Several supervised machine learning algorithms—including Support Vector Machine, Random Forest, and XGBoost—are evaluated, with LightGBM selected as the final classifier due to its computational efficiency and scalable leaf-wise boosting mechanism. Experimental results demonstrate that the framework delivers spatially consistent and accurate land-cover maps while maintaining lower computational complexity compared to deep learning-based alternatives. The modular design further enables seamless extension toward time-series land analysis and automated environmental monitoring workflows.