Document Type : Original Article
Authors
1
Dept of Photogrammetry and Remote Sensing, Faculty of Geomatics Engineering, K.N. Toosi University of Technology
2
Dept. of Photogrammetry and Remote Sensing, Faculty of Geomatics Engineering, K.N. Toosi University of Technology, Tehran, Iran
3
Department of Geomatics Engineering, Faculty of Civil Engineering, Tabriz University, Tabriz, Iran
4
School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran, Tehran, Iran
10.22059/eoge.2026.409826.1210
Abstract
Accurate segmentation of orthoimages remains a significant challenge in remote sensing, particularly within complex urban environments and low-quality datasets. This study introduces PSO-SAM, a novel hybrid framework that enhances the Segment Anything Model (SAM) by automatically tuning key input parameters—namely points_per_side and pred_iou_thresh—using Particle Swarm Optimization (PSO). Evaluated on a 2013 low-resolution orthoimages of Yazd, Iran, PSO-SAM demonstrates substantial improvements over the baseline SAM in all major segmentation metrics: IoU (0.1882, +65.5%), Precision (0.2705, +17.8%), Dice Coefficient (0.3168, +55.2%), Recall (0.3821, +108%), and F1-Score (0.3168, +55.2%). Although the achieved absolute segmentation metrics remain modest due to the challenging characteristics of the dataset and zero-shot limitations of SAM, the proposed optimization framework consistently improves segmentation quality across all evaluation criteria, and absolute metrics are reported to provide balanced interpretation. The sharp rise in Recall highlights PSO-SAM’s capacity to reduce false negatives and detect geographic features more comprehensively. To the best of our knowledge, limited studies have explored the integration of SAM with metaheuristic optimization strategies for urban aerial image segmentation. While absolute metric values remain lower than state-of-the-art benchmarks—primarily due to data limitations and the sensitivity of PSO—the results emphasize the potential of PSO-SAM for automated segmentation in constrained remote sensing scenarios. Future directions include improving data quality, exploring advanced optimization strategies, and leveraging higher-resolution or multispectral datasets to close the performance gap.
Keywords
Main Subjects
)