From Meta SAM to ArcGIS: A Comparative Analysis of Image Segmentation Methods for Monitoring Refugee Camp Transitions

This article presents a comprehensive evaluation of image segmentation methods for monitoring morphological changes in refugee camps, comparing five distinct approaches: ESRI Landviewer clustering, K-means clustering, U-Net segmentation, Meta’s Segment Anything Model (SAM) and ArcGIS segmentation. Using high-resolution satellite imagery from Al-Azraq refugee camp in Jordan (2014–2023) as a case study, this research systematically assesses each method’s performance in detecting and quantifying settlement pattern changes. The evaluation framework incorporates multiple validation metrics, including overall accuracy, the Kappa coefficient, F1-score and computational efficiency. The results demonstrate that ArcGIS’s ISO clustering and classification approach achieves superior performance, with 99% overall accuracy and a Kappa coefficient of 0.95, significantly outperforming the other tested methods. While Meta SAM shows promise in object detection, its performance degrades with aerial imagery, achieving only 75% accuracy in settlement pattern recognition. The study establishes specific parameter optimization guidelines for humanitarian contexts, with spectral detail values of 3.0–7.0 and spatial detail values of 14.0–18.0, yielding optimal results for refugee settlement analysis. These findings provide crucial methodological guidance for monitoring refugee settlement evolution and transition, contributing to more effective humanitarian response planning and settlement management through integrating remote sensing and machine learning technologies.