Segmentation Quality Analysis Using Segment Anything Model in Very High-Resolution Imagery

With the widespread use of artificial intelligence (AI) in many fields, automatic information extraction from satellite images is among the trending topics. Segment Anything Model (SAM) is one of the critical developments that has become very popular in this field. In this study, it is aimed to apply the SAM method to high-resolution images with heterogeneous classes and to investigate the quality of the object boundaries obtained. As the application area of the study, the Pléiades satellite image of the selected area in Bergama district of Izmir was used. The region has coniferous forests, buildings, roads, fruit trees, and bare land. The boundaries of the obtained objects were analyzed with Over-segmentation (OS), Under-segmentation (US), Area fit index (AFI), and Quality rate (QR), which are parameters that measure segmentation quality. In the numerical analysis, the results obtained with SAM were examined comparatively with the results of the MRS algorithm, which is one of the methods on which the OBIA algorithm is based. In this context, the SAM algorithm showed superiority in several parameters used to measure segmentation quality compared to the building class's multi-resolution segmentation (MRS) algorithm. However, for all parameters, the MRS algorithm was superior to SAM for the tree class of the selected test samples. In this way, a comparative study is presented using SAM and the widespread use of AI in different band combinations in high-resolution satellite images.