Refugee Camp Population Estimates Using Automated Feature Extraction

Brandon Green and Justine I. Blanford

Proceedings of the 53rd Hawaii International Conference on System Sciences, 2020


There is a growing trend in the use of aerial and satellite images to derive estimates of displaced populations in camps. High-resolution satellite imagery can be used to map physical structures in refugee and IDP camps, including changes to the number and type of these structures over time. Manual and automated feature extraction are two methods that can be used to map physical structures in refugee and IDP camps to support population estimates and geospatial analysis. Population estimates can be calculated by multiplying the number of dwellings by the estimated number of people per building, by multiplying the rooftop areas by the estimated average number of people per covered area, or by dividing the rooftop area by the estimated average covered area per person.

The authors of this paper develop a toolkit and workflow that can be used to automatically calculate estimates of displaced populations in camps based on feature information derived from an established automated extraction method. For the purpose of this study, the Rohingya refugee crisis was used, focusing on areas in and around existing refugee communities in two main refugee settlements, Kutupalong and Nayapara, in Bangladesh. Population estimates for each of the refugee camps were determined by: (a) identifying building features; and then using these features to (b) estimate the camp population based on the total area of the building features and UNHCR ‘covered area per person’ statistics. Accuracy of population estimates was determined by comparing the population estimates from the tool with those recorded by UNHCR for each camp.

This study demonstrates the potential scalable and transferable benefits of automated feature extraction methods, as the toolkit functioned as designed. A benefit of this method is the average processing time for each camp was 30 minutes compared to hours using manual extraction as demonstrated in other studies. However, the accuracy of automated tools using automated feature extraction methods rely on well-defined classifier definition files (used to classify pixels or group of pixels into different roof types and non-building features based on their spectral, textual or spatial properties). This study highlights the difficulty of developing well-defined classifier definition files that are geographically and temporally transferable.