Refugee camps often consist of more than 10,000 shelters and may need to be re-analysed several times in order to understand their evolution. Manual analysis typically leads to very high-quality output, but is very time-consuming. We have therefore worked with region-based convolutional neural networks to improve detection of new shelters in refugee camps, taking into account prior knowledge regarding the positions of existing shelters. The results were promising and the data pipeline created by our summer students has now been adapted and put to use by the UNOSAT experts. The retrained model yielded an average precision/recall score of roughly 80% and reduced the time needed for the task by a factor of 200 in some areas.

More recently, we also addressed the challenge of simulating synthetic high-resolution satellite images. High-resolution satellite imagery is often licensed in such a way that makes it difficult to share it across UN partners and academic organisations. This reduces the amount of image data available to train deep-learning models, thus hampering research in this area. We have developed a generative adversarial network (GAN) that is capable of generating realistic satellite images of refugee camps images. Our tool was initially based on a progressive GAN approach developed by NVIDIA. We have now developed this further, such that it can combine multiple simulated images into a cohesive larger image of roughly 5 million pixels.

Several other lines of investigation — all related to AI technologies — are also being pursued within the scope of this project.