Nowadays, modern space missions continuously collect information about the earth surface that corresponds to massive amounts of data. The multitude of Earth Observation (EO) systems allows the acquisition of data via different sensors (e.g., optical, radar, LiDAR) at different spatial and temporal resolutions, with diverse spectral characteristics. This huge and diverse volume of information opens up new opportunities to better understand and monitor agricultural, natural and anthropized spaces at different scales.
In this context, data-intensive methodologies such as machine and deep learning approaches are demonstrating their value, as they already did in several domains dealing with signal data. Multiple data science challenges were already addressed using satellite imagery (e.g., building footprints, road networks, iceberg detection) but crucial open questions remain unsolved (e.g., biodiversity monitoring, urban mapping, deforestation tracking and food risk prevention, triaging disaster zones). We are at the beginning of a new era for the analysis of Earth Observation data (EOD) where one of the main questions is how to leverage the complementarity and the diversity of the information collected by the different available observation systems, in order to answer important societal challenges and monitor changes on the Earth Surface.
The MDL4EO team (Machine and Deep Learning for Earth Observation) at the UMR TETIS (Montpellier, France) has the objective to scientifically contribute to this new era providing AI methods and algorithms able to extract valuable knowledge from massive heterogeneous Earth Observation Data.
In this tutorial, we will discuss in detail the main research questions addressed by the MDL4EO team :
Recent years have witnessed a Cambrian explosion of tools and techniques able to tackle problems that were only solvable by humans up to a few years ago; deep learning in particular is accumulating astounding successes at a breakneck pace in both research and applications: from helping in recovering photos by their descriptions on devices used by billions of people, to providing tools for investigating the depths of the visible universe. It is then unfortunate that these very models are utterly unscrutable and inaccessible to human understanding.
While in many cases the difficulty of understanding these models does not matter, in some very specific contexts it creates problems that are important and hard to solve. Big and small companies are, in fact, investing in these technologies and deploying them in contexts that directly impact human well-being such as loan applications, candidate selection for job offers, and evaluating the chance of re-offending for people who commited crimes. In all these cases using unscrutable models poses difficult etical issues related to the risk of discrimination of people belonging to protected groups.
In absence of techniques allowing to solve the problem by explaining the decisions of these models, the fair ML literature focused on approaches based on the construction of non-discriminating models. In this context, neural networks provide both challenges and opportunities. In this tutorial I will contextualize the problem, show how there exists many different (and contrasting) definitions of fairness, and introduce some of the state-of-the-art approaches in this field. I will focus in particular on methods targeting neural networks, specifically on methods that constrain the representation learnt by the network to be fairer with respect to given sensible attributes.