Evaluation of image-based land use classification with multi-feature fusion approach

Abstract

Traditional land use and land cover (LULC) classification approaches using spectral bands display persistent limitations in distinguishing similar landcover types, particularly within heterogeneous urban–rural transition zones. This research examines the effectiveness of multi-feature fusion for improving classification accuracy by systematically evaluating combinations of spectral indices with machine learning classifiers. Sentinel-2 imagery of Kaduwela, Sri Lanka, was utilized, focusing on six LULC categories: water, forest, vegetation, roads, buildings, and rock exposure. The methodology involved an initial evaluation of baseline RGB performance using Support Vector Machine (SVM), Random Forest (RF), and Decision Tree (DT) classifiers. Subsequently, 27 spectral indices were generated and integrated with RGB to form a 30-dimensional feature set, followed by Principal Component Analysis (PCA) to reduce dimensionality, retaining five components that explained over 95% of the variance for subsequent classification. The RGB feature set yielded accuracies of 39.22% (SVM), 44.19% (DT), and 50.45% (RF); integration of RGB with 27 indices improved accuracies to 53.22% (SVM), 52.01% (DT), and 52.94% (RF); while the PCA-reduced feature set provided 54.70% (SVM), 46.18% (DT), and 49.34% (RF). The findings highlight that in heterogeneous and urban–rural interfaces, PCA-based reduction of spectral indices has improved SVM classification and reduced computational load. For future studies, further hyperparameter tuning could enhance accuracy.