The first AI4EO Symposium by International Future Lab in Munich
On 13th October, researchers and industry experts interested in AI applications for Earth Observation (EO) gathered in Ottobrunn, a Technical Univesity of Munich (TUM) Campus near Munich for the first AI4EO Symposium.
Given that my abstract for the paper on predicting urban vitality from satellite data (paper website) was selected for an oral presentation at the Symposium, I could not miss coming from Oxford for this exciting event.
Prof. Xiaoxiang Zhu gave the opening speech in which she presented the works of her teams at TUM and German Aerospace Center (DLR). It was humbling to learn about the breadth and depth of the work done by the lab: from using social media images for building function classification to EarthNets, an amazing collection and analysis of EO datasets published openly along with AI models applicable to them.
Next, Dr Rune Floberghagen EO Mission Manager from European Space Agency (ESA) presented his agency’s views on the current state of AI in EO. Uncertainty, fairness, trustworthiness and other aspects of Responsible AI were accented, along with the fact that the applications of AI in most subfields now follow a top-down approach (that is to say, AI became available, how can we incorporate it best into our practices?) except for climate science (where the situation is that we have an issue and ask how can we solve it — and then we find that AI is the solution where previous approaches could not do it).
During the Symposium, there were two keynote talks devoted to the uncertainty of AI models in EO (one by Prof. Yuanyuan Wang and another by Prof. Lorenzo Buzzone). Both highlighted the need for models to output how certain they are and explained the sources of uncertainty (from labels to the data to the model itself). This again is tightly linked to the need for Explainable AI.
Finally, Prof. Jonathan Bamber and Prof. Richard Bamler led a session of oral presentations on Climate while Prof. Lichao Mou led the session on Novel Methods and Data, in which I presented, too.
I will highlight only a couple of (lightning) talk ideas that impressed me. We can build a Gini Coefficient for urban green space inequality. (Contrastive) Self-Supervised Learning (SSL) is a very popular approach at the moment as it tries to tackle the limited availability of labelled data. Combining NLP models with computer vision ones will enable visual question answering. There are platforms, such as ExtremeEarth and AI4Copernicus that aim to enable the extraction of information and knowledge from big Copernicus data using deep learning techniques. We can quantify health risks with AI4EO (e.g., Dengue). Graph Neural Networks (GNNs) found their application in Atmospheric studies for in-painting missing remote sensing data. There are advanced video games for data augmentation (labelling) via gaming.
