|
Meta publicly released the Segment Anything Model (SAM), an AI model specialized for image segmentation purposes, in 2023. Then, based on this scheme, there has been a lot of research on applying it to object recognition and extraction problems in various applications. There is also active research on derivation techniques to improve the performance and accuracy of SAM. SAM supports zero-shot or few-shot generalization, which does not require the collection of its segmentation data and does not require the model to be labeled for the use case. This research is a case study of SAM application using a fusion of high-resolution optical and SAR imagery as input. The images used in this study are KOMPSAT-3/3A optical images and KOMPSAT-5 Synthetic Aperture Radar (SAR) images, and the purpose of the application is water object extraction. The input images are fused by applying the wavelet technique. The study area is Boryeong City in Chungcheongnam-do, west coast area in the Korean Peninsula. The results obtained by applying SAM to this data were compared with the digital map of the river boundary, lake, or reservoir of the National Geographic Information Center in Korea and the OpenStreetMap (OSM) water objects in the same study area for quantitative accuracy comparison analysis. The Weight mean Intersection over Union (WmIoU) technique was used as an accuracy evaluation method. The experimental results show a significant improvement in accuracy when fused images are applied compared to optical images. As a result of this study, we expect that SAM, specializing in natural object extraction for multi-sensor image data, will be developed in the future and can be applied to solve many current problems.
|
