Abstract
Abstract. Image segmentation, is of utmost importance in the disciplines of digital image processing, particularly remote sensing and computer vision, has seen an increasing demand for precise and efficient algorithms. This study focuses to conduct a comparative exploration of the segmentation capabilities of two sophisticated techniques namely, Multiresolution Segmentation (MRS) and Segment Anything Model (SAM), leveraging the high-resolution WorldView-3 (WV-3) satellite image. MRS adopts a hierarchical methodology, segmenting an image into various scales while retaining a profound understanding of its structure. Conversely, SAM employs a deep learning algorithm, prioritizing segment creation based on conceptual pixel similarity, irrespective of spatial adjacency. The WV-3 image, featuring diverse land cover elements like agricultural parcels, industrial structures, roads, red roofs, single trees, and water bodies, serves as the basis for assessing segmentation quality. Both methods are applied to the image, and their outcomes are individually evaluated against manually generated polygonal land use/cover objects. Segmentation quality metrics are employed for assessment. Results reveal MRS effectively preserves fine details and entity delineation, while SAM excels in capturing contextually similar regions. MRS outperforms SAM with a negligible discrepancy, yet SAM demonstrates superiority in the red roof object, achieving an Intersection over Union (IoU) value of 0.70 compared to MRS’s 0.49. MRS tends to generate numerous segments for an item, while SAM produces only one segment. Nevertheless, it is important to recognise that both algorithms have specific constraints in particular scenarios, such as excessive segmentation in areas with abundant texture or inadequate segmentation in areas with slight changes.