Abstract
With the global refugee crisis at a historic high, there is a growing need to assess the impact of refugee settlements on their hosting countries and surrounding environments. Because fires are an important land management practice in smallholder agriculture in sub-Saharan Africa, burned area (BA) mappings can help provide information about the impacts of land management practices on local environments. However, a lack of BA ground-truth data in much of sub-Saharan Africa limits the use of highly scalable deep learning (DL) techniques for such BA mappings. In this work, we propose a scalable transfer learning approach to study BA dynamics in areas with little to no ground-truth data such as the West Nile region in Northern Uganda. We train a deep learning model on BA ground-truth data in Portugal and propose the application of that model on refugee-hosting districts in West Nile between 2015 and 2020. By comparing the district-level BA dynamic with the wider West Nile region, we aim to add understanding of the land management impacts of refugee settlements on their surrounding environments.
Introduction
In sub-Saharan Africa, fire is a critical tool for land management in smallholder agriculture. Understanding "burned area" (BA) dynamics is essential for assessing environmental health and the impact of human settlements. However, researchers face two major hurdles: a severe lack of ground-truth data in the region and the coarse resolution of existing global products. For instance, the MODIS-based MCD64A1 product has a 500m spatial resolution, which is often too coarse to detect the small-scale "slash and burn" practices common in Uganda, potentially missing up to 70% of the total burned area.
Methodology
To overcome the data scarcity in Northern Uganda, this study utilizes a transfer learning pipeline. The core strategy involves training a deep learning model on high-quality wildfire data from Portugal and transferring that knowledge to the Ugandan context.
The team employed a U-Net architecture, utilizing a ResNet-18 encoder for feature extraction and a PSPNet (Pyramid Scene Parsing Network) decoder for map reconstruction. The model processes Sentinel-2 satellite imagery, specifically focusing on the Near-Infrared (8A), Green (3), and Shortwave-Infrared (12) bands, which are highly effective at highlighting burned vegetation.
Datasets and Training
The model was trained on a robust dataset of 2,704 patches from Portugal (2016), chosen for its high-quality ground-truth labels and semi-arid climatic similarities to Northern Uganda. For inference in Uganda, the model analyzed Sentinel-2 Level-1C data from 2015–2020 across districts like Arua, Yumbe, and Moyo, which host major refugee settlements such as Bidi Bidi and Rhino Camp.
Key Results and Insights
On the Portuguese test set, the model achieved an Intersection-over-Union (IoU) of 0.559 and a Dice score of 0.661. When applied to the West Nile region, the model detected a significantly larger amount of burned area compared to the MODIS baseline. Visual inspections confirmed the model's ability to identify fine-grained burn scars that standard global products missed entirely.
A key insight from this work is that high-quality labels from one geographic region can effectively bridge the data gap in another. Furthermore, the model demonstrated resilience to label noise, sometimes predicting fire locations that manual ground-truth labels had overlooked.
Conclusion and Future Work
This research provides a scalable pathway to study the environmental impacts of refugee settlements. By using 20m resolution Sentinel-2 data, the model offers a much clearer picture of land management intensity than previously possible. Future work will address limitations such as domain shift and atmospheric correction, with the ultimate goal of helping humanitarian organizations plan more sustainable settlements through a better understanding of the refugee-environment nexus.
Summary updated