Humanitarian AI applications for satellite imagery

Project goal

This project is making use of expertise in artificial intelligence (AI) technologies at CERN to support a UN agency. Specifically, we are working on AI approaches to help improve object recognition in the satellite imagery created to support humanitarian interventions. Such satellite imagery plays a vital role in helping humanitarian organisations plan and coordinate responses to natural disasters, population migrations, and conflicts.

R&D topic
Applications in other disciplines
Project coordinator(s)
Sofia Vallecorsa
Team members
Suren Thapa
Collaborator liaison(s)
Lars Bromley, Edoardo Nemni

Collaborators

Project background

Since 2002, CERN has hosted UNOSAT, the Operational Satellite Applications Programme of UNITAR (The United Nations Institute for Training and Research) on the laboratory’s premises. UNOSAT acquires and processes satellite data to produce and deliver information, analysis, and observations to be used by the UN or national entities for emergency response, to assess the impact of a disaster or a conflict, or to plan sustainable development in the face of climate change.

At the heart of this project lies the idea of developing machine-learning techniques that can help speed up analysis of satellite imagery. For example, predicting and understanding the movement of displaced persons by identifying refugee shelters can be a long, labour-intensive task. This project is working to develop machine-learning techniques that could greatly reduce the amount of time needed to complete such tasks.

Recent progress

In 2020, we focused on the challenge of simulating synthetic high-resolution satellite images. High-resolution satellite imagery is often licensed in a way that makes it difficult to share it across UN partners and academic organisations. This reduces the amount of image data available for training 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. 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. The new model is built on a multi-network architecture combining several auto-encoders; their output is used to condition the image-generation step and to ensure each new image is consistent with previous ones. This method was tested on satellite images of a flooded area in Myanmar.

Next steps

Next year will be dedicated to the optimisation of the progressive GAN model. In particular, we will implement a distributed approach for training our network in parallel across multiple nodes. This should help us to reduce training time for the model and increase the maximum image size.

Publications

    N. Lacroix, T. Aliyev, L. Bromley: Automated Shelter Recognition in Refugee Camps. CERN openlab Summer Student Report. Published on ZENODO, 2019. cern.ch/go/v6rn

Presentations

    Y. Boget, ProGAN on Satellite images (15 August). Presented at CERN openlab summer student lightning talk session, Geneva, 2019. cern.ch/go/P6NV
    Y. Boget, S. Vallecorsa, Deep Learning for Satellite Imagery (24 September). Presented at IXPUG Annual Conference, Geneva, 2019. cern.ch/go/m9n6