Object-based image analysis for fine-grained classification and dynamic change monitoring of grassland and shrubland in karst regions using UAV RGB imagery

Abstract

Fine-grained identification of grassland and shrubland in Karst regions remains a challenge for traditional satellite remote sensing due to fragmented landscapes and high spectral similarities between vegetation types. This study evaluates a cost-effective and deployable vegetation monitoring solution using high-resolution visible-light (RGB) imagery acquired by DJI Phantom 4 Pro UAVs in 2021 and 2022. Operating within an Object-Based Image Analysis (OBIA) framework, we systematically optimized multi-scale segmentation parameters and evaluated the classification performance of three supervised learning algorithms: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Decision Tree (DT). Furthermore, we analyzed the spatiotemporal patterns and dynamic transitions of primary land cover types. Our results indicate that a parameter set comprising a scale of 100, shape of 0.3, and compactness of 0.3 constitutes the optimal segmentation configuration, effectively delineating land cover boundaries while maintaining object homogeneity. Both KNN and DT significantly outperformed SVM; KNN achieved the highest overall accuracy (OA = 0.8764 in 2021; 0.8754 in 2022) and superior inter-annual stability, whereas DT demonstrated the highest adaptability to varying annual imaging conditions. From 2021 to 2022, shrubland coverage increased from 32.15% to 38.42%, expanding toward the central region, while grassland coverage decreased from 28.72% to 24.16%. The continuity of vegetation restoration patches improved, and artificial land remained stable. This study validates the feasibility of utilizing lightweight RGB-equipped UAVs for precision monitoring of Karst ecosystems, providing a pragmatic technical benchmark for regional rocky desertification control and ecological assessment.