Assessment of the Segment Anything Model’s Applicability to Coastal Debris Segmentation: A Comparative Analysis of Prompt input Methods

Abstract: Coastal debris poses serious threats to marine ecosystems and fisheries, increasing the demand for effective detection technologies. However, the irregular shapes, low color contrast, and complex backgrounds of such debris limit the generalization capability of supervised segmentation models. To address this limitation, this study evaluates the applicability of the Segment Anything Model (SAM), which performs object segmentation based solely on prompt inputs without requiring additional training. The segmentation performance of SAM was quantitatively assessed under various conditions by comparing Ground Truth (GT)-based prompts with automatically generated prompts. Although automatic prompts generally yielded lower performance than GT-based ones, relatively high accuracy was observed for objects with distinct color contrast or simple backgrounds. In particular, the automatic point-based prompt achieved performance comparable to GT for certain object categories. These results suggest that, with well-designed prompts, SAM can produce a certain level of segmentation performance even in complex coastal environments. This implies that SAM may be useful for preliminary object detection in previously unseen regions where labeled data are unavailable. Furthermore, optimizing prompt design in combination with post-processing strategies could improve its practical applicability in coastal areas lacking sufficient training data.