Comparative Analysis of Segmentation Parameters in Object-Based Image Classification for Land Cover/Land Use in Sumedang Regency

Abstract

Information on land cover and land use plays an important role in biophysical regional analysis, spatial planning, resource management, and the formulation of sustainable development policies. To support these needs, accurate remote sensing image classification is crucial. The Object-Based Image Analysis (OBIA) approach is considered superior to pixel-based classification, as it provides higher accuracy and minimizes the salt-and-pepper effect. The success of object-based classification is influenced by the segmentation method employed. In this study, two segmentation approaches, namely Original Multiresolution Segmentation (OMN) and Region Grow on Object (RGO), were examined based on combinations of segmentation parameters and evaluated for accuracy using the Random Forest (RF) classification algorithm. The segmentation results indicate that the OMN approach produces smaller and more detailed objects, though they tend to be fragmented, whereas the RGO approach generates larger and more generalized objects with greater spatial stability. Based on object-based classification using RGO segmentation with a scale parameter of 0.5, seven land cover and land use classes were identified, with three dominant categories: plantations (50,393 ha), bare land (29,658 ha), and rice fields (27,092 ha). The classification accuracy of RGO was consistently higher than OMN across all parameter configurations, with the rice field class showing close alignment with official BPS data from 2024, which recorded an area of 30,038 ha. These findings demonstrate that the RGO approach is more effective in producing representative segmentation and classification for land use mapping.