Enhancing Land Use/Land Cover Analysis with Sentinel-2 Bands: Comparative Evaluation of Classification Algorithms and Dimensionality Reduction for Improved Accuracy Assessment

Abstract

Accurately classifying land use and land cover (LULC) is crucial for understanding Earth's dynamics under human influence. This study proposes a novel approach to assess LULC classification accuracy using Sentinel-2 data. Authors have compared traditional and Principal Component Analysis (PCA)-based approaches for Maximum Likelihood Classification, Random Forest, and Support Vector Machine (SVM) classifiers. Four key classes (agricultural land, water bodies, built-up areas, wastelands) are classified using supervised learning. Accuracy is evaluated using producer, user, overall accuracy, and kappa coefficient. Our findings reveal superior accuracy with PCA-SVM compared to other methods. PCA effectively reduces data redundancy, extracting essential spectral information. This study highlights the value of combining PCA with SVM for LULC classification, empowering policymakers with enhanced decision-making tools and fostering informed policy development.