Investigating the capability of UAV imagery for AI-assisted mapping of Refugee Camps in East Africa

Abstract

Refugee camps and informal settlements provide accommodation to some of the most vulnerable populations, with many of them located in Sub-Saharan Africa. Many of these settlements lack up-to-date geoinformation that we take for granted in developed world. Having up-to-date maps on their dimension, spatial layout is important. They are essential tools for assisting administration tasks such as crisis intervention, infrastructure development, and population estimates which encourage economic productivity. In the OpenStreetMap ecosystem, there is a disparity between built-up being digitised in the developed and the developing areas. This data inequality are results of multiple reasons ranging from a lack of commercial interest to knowledge gaps in data contributors and such disparity can be reduced with the help of assisted mapping technology. Very High Resolution remote sensing imagery and Machine Learning based methods can exploit the textural, spectral, and morphological characteristics and are commonly used to extract information from these complex environments. In particular recent advances in Deep Learning based Computer Vision have achieved significant results. This study is connected to a larger initiative to open-source the AI assisted mapping platform in the current Humanitarian OpenStreetMap Team's ecosystem, to investigate the capabilities of applying Deep Learning for building footprint delineation in refugee camps based on open-data Unmaned Aerial Vehicle (UAV) imagery from partner organisation OpeAerialMap. The objective of this study is to test the U-Net and several variations of the architectures' performance for building footprint segmentation, The performance of the different Deep Learning models on datasets of various complexity were collected. A comparison of the models' responses using class-based accuracy assessments metrics allows detail evaluation into how the different architectures and experiment setup respond to data quality. Given the computation and resources constraint of this project, the result suggests that increase in architectural depths corresponds with increase in precision. Models that were initialised on pre-trained weights from ImageNet could reduce recall. Lastly, to our surprise, the transferability of a competition winning network trained on similar resolution but on formal building performs worse than many models trained from scratch. This study showcased the ability to use Deep Learning semantic segmentation to perform building footprint delineation in complex humanitarian applications. Having increased access to open-data Very High Resolution UAV imagery from the OpenAerialMap initiative is an advantage to building AI-assisted humanitarian mapping. The study demonstrated a careful and rigourous approach to model evaluation. Yet, the variation of the study results not only emphasised the complexity of Deep Learning based methods, but also indicate the direction for further investigation that would be justifiable when further resources becomes available.