Abstract: Monitoring basic services in the hyper-dense Rohingya refugee camps of Cox’s Bazar, Bangladesh, presents a significant logistical challenge. This study introduces a semi-supervised AI framework utilizing sub-meter satellite imagery and the Segment Anything Model (SAM) to track Water, Sanitation, and Hygiene (WASH) accessibility. By integrating high-resolution shelter detection with spatial accessibility modeling, the research identifies a critical decline in service levels between 2022 and 2025. The findings highlight the impact of population growth and infrastructure degradation, particularly affecting women and girls.
Introduction
In humanitarian crises, effective management is predicated on accurate measurement. The Rohingya refugee camps are among the most densely populated areas globally, making the monitoring of basic services like water pumps and latrines difficult. Traditional ground surveys are slow and often outdated by the time they are completed. While remote sensing offers a scalable alternative, the camp's irregular morphology and high density often cause standard computer vision models to fail. This paper bridges this gap by creating a robust framework for mapping shelters and calculating real-world accessibility to life-sustaining infrastructure.
Methods
The researchers employed a two-stage methodological framework. The first stage involves shelter detection using a semi-supervised "student-teacher" paradigm. A teacher model provides initial predictions on unlabeled satellite data, which are then refined using the Segment Anything Model (SAM) to sharpen boundaries and filter out noise such as shadows and vegetation. A student model is then trained on these refined labels.
The second stage utilizes the Enhanced Two-Step Floating Catchment Area (E2SFCA) method for accessibility modeling. This model considers supply (number of facilities), demand (estimated population based on detected shelter area), and friction (actual walking distances along footpath networks). Crucially, the analysis includes gender-disaggregated modeling to account for safety barriers faced by women.
Key Results and Insights
The proposed segmentation framework achieved an F1-score of 76.4%, significantly outperforming standard models like U-Net (70.3%) and FCN (71.5%). This precision allowed for a detailed analysis of accessibility trends. The study found that the average number of people served by a single WASH facility rose from 25 in 2022 to 29.4 in 2025. Water pumps are the most strained resource, now serving an average of 67.1 people per pump.
A key insight is the "crowding out" effect: while the physical footprint of the camps has stabilized since 2020, the population has continued to grow, leading to a "vertical" increase in density. Furthermore, the study identifies a significant "safety gap." When accounting for a 25% reduction in effective capacity for women due to safety concerns at communal facilities, accessibility scores for females plummeted compared to males.
Discussion and Conclusion
The decline in accessibility is attributed to a combination of population growth and infrastructure degradation, likely exacerbated by funding cuts. While the model uses shelter area as a proxy for population, which may not account for multi-story structures, its accessibility scores show a strong correlation (Spearman $\rho = -0.71$) with independent ground-truth survey data.
Future work will involve integrating real-time maintenance data and using higher-resolution temporal data to track camp changes following disasters like fires or floods. This high-resolution mapping enables humanitarian agencies to identify micro-areas of extreme neglect and optimize resource allocation, supporting more effective and equitable aid delivery in one of the world's most challenging environments.
Summary updated