Multispectral UAV-based LULC mapping performance improvement by integrating precise NDSM data and machine learning algorithms

Abstract The advancement of sensor technologies has enabled the production of high-resolution Land Use and Land Cover (LULC) maps, enhancing change detection in land surface dynamics. UAV-derived high-resolution data improves the performance of segmentation and classification by highlighting object heterogeneity. This study focuses on performance improvement in multispectral (MS) UAV-based LULC mapping by incorporating high-accuracy Normalized Digital Surface Model (NDSM) data along with multiple indices from literature in a test area where multiple terrain classes with significant elevation heterogeneity (up to 36 m) exist. The primary objective is identifying the most effective dataset and classification algorithm by evaluating NDSM separately in segmentation and classification. To that end, Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost) machine learning algorithms were used on four datasets created from spectral, index, geometry, texture and NDSM feature groups obtained from object-based image analysis (OBIA) of the produced MS UAV orthomosaic and the LULC mapping performance were evaluated by accuracy metrics mostly preferred in the literature. The results showed that including NDSM in the dataset improved the overall accuracy of all classifiers by 4% to 7% compared to the other datasets. The highest overall accuracy (94.65%) was achieved using XGBoost on the dataset including NDSM. Subsequently, a comprehensive class-based analysis of all influential features contributing to this outcome was conducted utilizing the SHapley Additive exPlanations (SHAP) algorithm. The results showed that NDSM-derived elevation data had the strongest impact on class separability, enhancing thematic map accuracy.