A new study involving the UGR shows that deep generative vision models can synthesise highly realistic satellite images that can be used to illustrate future climate-related events such as floods, sea-ice retreat, and reforestation.

Natalia Díaz, a UGR researcher at the Andalusian Inter-University Institute for Data Science and Computational Intelligence (DaSCI), participated in the project in collaboration with researchers from centres and foundations in the United States (including MIT), Canada, Germany and the United Kingdom.

To tackle their research problem, the team used a generative adversarial network (pix2pixHD) to create synthetic satellite images of future flooding, as well as the effects of positive actions such as reforestation. Although the model is capable of generating realistic images, it tends to ‘hallucinate’ floods in the wrong areas. Their proposed solution to this problem combines deep learning with segmentation maps generated by physics-based flood models. This improved approach outperforms not only the pure deep learning model, but also manual solutions, by significantly reducing prediction errors and improving the reliability of the images.

The research team evaluated the generalisability of their method using multiple remote sensing datasets and climate-related events, including reforestation and melting Arctic sea ice. In addition, they made their code, new metrics, and an extensive dataset available to the scientific community, including more than 30,000 labelled HD image triplets for segmentation-guided image-to-image translation, which is equivalent to 5.5 million images at 128×128 pixels.

This project marks a crucial step towards the creation of reliable and realistic visual tools for communicating the impacts of climate change, establishing further grounds for greater collaboration between the fields of physics-based modelling and deep learning.

The Andalusian Inter-University Institute for Data Science and Computational Intelligence (DaSCI) is a collaborative organisation that is managed jointly by the universities of Granada, Jaén and Córdoba. Its mission is to conduct advanced research and training in the field of artificial intelligence, with a particular focus on data science and computational intelligence. The Institute brings together preeminent researchers working on joint projects and promotes the development and application of innovative technologies in diverse sectors. With a view to becoming a leading organisation in its field, the DaSCI fosters the transfer of scientific knowledge to the socio-economic environment, thus contributing to technological progress and the digitisation of industry.