The United Nations Institute for Training and Research (UNITAR) provides innovative learning solutions to individuals, organizations and institutions to enhance global decision-making and support country-level action for shaping a better future.
Description of Activities on AI
UNITAR has multiple activities of interest within the areas of artificial intelligence and satellite imagery analysis that will further progress to the Sustainable Development Goals (SDGs). Since 2000, UNITAR has developed its Operational Satellite Applications Programme (UNOSAT), which has focused on satellite imagery analysis in support of UN-related operations within the areas of disaster response, human rights, security, and development, and capacity development within these areas. This has allowed UNOSAT to develop decades of in-house expertise in satellite imagery analyses on issues vital to UN operations. Just as important, UNOSAT has for many years produced a robust collection of vector datasets with its analysis results that prove extremely useful as training data for AI and related development. Much of the success of UNOSAT computing activities result from excellent major collaborations with CERN Openlab and UN Global Pulse, UNICEF Innovation, UNHCR Innovation, ITU, and other partners.
UNOSAT projects in development in these thematic areas include the development of specific algorithms as well as tool development for analyzing satellite imagery across the UN system. These different aspects are described in greater detail below.
Project 1: Algorithm development
With decades of experience analyzing satellite imagery and mapping natural disasters, refugee settlements, conflict, and related issues UNOSAT has emerged as a primary partner for organizations wishing to explore AI for humanitarian applications. In addition, large amounts of UNOSAT analyses are publicly available in vector format both from its’ website and the Humanitarian Data Exchange, and this has proven valuable for organizations seeking training data for AI development. In turn, multiple such organizations have reached out to UNOSAT for guidance on imagery analysis, allowing UNOSAT to learn a great deal about the state of the field. This partnership model has been very effective in particular with UN Global Pulse and the two organizations have collaborated extensively to develop algorithms that can identify and map refugee shelters in satellite imagery. Importantly, UNOSAT and Pulse have paid particular attention to the accuracy of outputs of these algorithms given the high-threshold of accuracy UNOSAT requires for its operations. This process and results were detailed in a 2018 academic paper by Pulse and UNOSAT. Currently, UNOSAT also developed an end-to-end pipeline where images of flood-prone areas are automatically downloaded from satellites covering the affected areas and processed by machine learning algorithms to shorten the time needed to deliver disaster maps to humanitarian organizations. The system would allow for near-real-time monitoring and surveillance. UNOSAT is also pursuing algorithm development for change detections which would help enable various analyses such as landslide detection and damage assessments. Finally, UNOSAT together with CERN and European Space Agency will engage in a challenge over the next several months to extract building footprints from satellite imagery.
Project 2: Tool development
Concurrently with algorithm development UNOSAT and Pulse are working to develop a cloud-based infrastructure for processing and analyzing satellite imagery using AI models. This tool is in an ‘early beta’ phase and has been shared with a few UN partners for testing. UNOSAT is providing feedback on usability from the perspective of an ‘expert GIS’ user. Eventually this tool is intended to provide access to large amounts of satellite imagery as well as AI-based analysis methods for the UN community to use without requiring any specialized hardware or software themselves.
Challenges and Opportunities
UNOSAT together with its partners has made quite impressive progress on developing AI methods for satellite imagery analysis, but additional funding is needed in order to scale up the efforts, which would likely achieve amazing results. The before mentioned training data coupled with in-house expertise and understanding of requirements are the basis for excellent opportunities in advancing AI for satellite imagery analysis.