.jpg)
In the realm of self-supervised learning (SSL) applied to satellite imagery, a key question arises: what information do the learned representations actually encode if they are not evaluated solely with concrete tasks? An innovative approach proposes probing these representations using physically consistent environmental variables, such as temperature, precipitation, or solar radiation, extracted from the ERA5 reanalysis. These variables are not part of the SSL training but maintain a causal relationship with the spectral reflectance and radar backscatter recorded by sensors like Sentinel-1 and Sentinel-2. By analyzing models such as DINO, MAE, and MoCo —trained under identical conditions— it is discovered that intrinsic metrics of the representation reveal differences that traditional benchmarks overlook. This finding is relevant for the development of geospatial foundation models, where the ability to capture environmental signals can improve tasks such as crop yield prediction, water resource monitoring, or climate risk management.
From a business perspective, integrating robust semantic representations with environmental data opens opportunities to build AI for businesses that transform Earth observation into operational decisions. For example, an artificial intelligence system that combines satellite imagery with climate variables can optimize agricultural planning or early fire detection. At Q2BSTUDIO, we develop custom applications that leverage these advances, whether through AWS and Azure cloud service pipelines to process large volumes of geospatial data, or through business intelligence services with Power BI to visualize correlations between SSL representations and environmental phenomena. Additionally, our cybersecurity expertise ensures that sensitive data remains protected throughout the entire model lifecycle.
The approach of evaluating SSL representations with environmental signals also reinforces the need for AI agents capable of reasoning about physical contexts. Instead of relying exclusively on human labels, these agents can exploit causal relationships inherent in the data, improving their generalization ability. Therefore, at Q2BSTUDIO we advocate for custom software that integrates these principles, facilitating the creation of robust solutions for sectors such as precision agriculture, logistics, or urban planning. The combination of self-supervised representations with physical variables not only increases model transparency but also paves the way toward more reliable AI systems aligned with the reality of the natural world.
