Artificial Intelligence
From Deep Learning to Generative AI for Earth Observation
Climate crisis is compelling us to better understand the complexity of our planet by extracting useful information quickly and on a global scale. In this context, Artificial Intelligence is a key element for making Earth Observation (EO) more timely, accurate, and automated.
Today, AI represents one of the most dynamic areas in the space sector, where it is beginning to be widely and concretely applied thanks to the convergence of several factors: more observational data, increased computing power, advances in algorithms, and the spread of scalable infrastructures.
MEEO also embraces this innovation, integrating established EO data science methods with advanced deep learning techniques, often combined in hybrid systems to maximize the performance and reliability of analyses.
1
Deep learning
The starting point of this innovation
Deep Learning (DL) is a machine learning technique that uses multiple layers of neural networks to learn complex tasks.
In the field of EO, it is now widely used for:
Object Detection
The model learns to locate and identify specific objects in satellite imagery.
Super Resolution
The model learns to reconstruct missing details in low-resolution satellite images increasing the native resolution, helping to overcome in a more realistic way the physical limitations of satellite sensors.
Semantic Segmentation
Assigns a specific class label or semantic meaning to each pixel in a satellite image.
MEEO application example
Lake Environmental Monitoring project
Satellite-derived Lake Surface Water Temperature (LSWT) is a key indicator of lake health, heatwaves, and long-term climate trends, but the daily record is often fragmented by cloud cover and acquisition gaps, especially for small lakes, where entire scenes can be missing for several consecutive days.
Here MEEO adapts self-supervised deep learning (Masked Autoencoders) to reconstruct cloud-free daily LSWT fields, learning spatio-temporal patterns from Earth Observation time series (ESA Lakes CCI) and inferring missing observations in a physically consistent way, without manually labelled data.
The result is a consistent, gap-free time series that scales from single-lake pilots to global, multi-lake monitoring, turning fragmented satellite records into reliable evidence for environmental decision-making.
2
Foundation Models
A paradigm shift
Unlike traditional deep learning approaches, Foundation Models do not require training from scratch because they come with a pretrained backbone that can be quickly fine-tuned to new specific downstream tasks, optimizing a range of activities such as classification, segmentation, and change detection.
In the field of Earth observation, foundation models are typically trained using multimodal data, including multispectral, SAR, and auxiliary geospatial layers.
MEEO application example
DVPS project
DVPS (Diversibus Viis Plurima Solvo) is an EU-funded research project (2025–2029) that aims to turn the design of multimodal foundation models into a rigorous, scalable scientific discipline. It develops methodologies for integrating hundreds of heterogeneous data modalities (far beyond conventional text, audio, and image inputs) into unified learning architectures, delivering AutoDVPS, an open-source toolkit for automated multimodal model design, pre-training, and fine-tuning, together with DVPS-FM, a general-purpose multimodal foundation model trained across diverse application domains.
Within the Geo-Intelligence domain, MEEO is a core project partner, bringing its expertise in Earth Observation data management and advanced geospatial analytics for agricultural applications, with a focus on multimodal data curation, preprocessing, and cross-modal alignment to deliver high-quality, ML-ready datasets. In Geo-Intelligence Use Case 3, MEEO integrates satellite imagery with complementary geospatial sources to support agricultural ecosystem monitoring, enabling downstream tasks such as crop type classification, vegetation state estimation, and large-scale ecosystem monitoring and ultimately data-driven decisions for climate-resilient and sustainable agriculture.
3
Generative Artificial Intelligence
The turning point
The introduction of Transformer architecture in 2017 marked a major breakthrough in artificial intelligence. These are deep learning models that compute how all parts of a given input (e.g., text or image) connect and influence each other, thereby understanding the global context simultaneously rather than sequentially.
This is the foundation on which generative AI is built.
At MEEO, we adopt this technology, trained on heterogeneous Earth observation and remote sensing data.
Agentic frameworks
A key research direction is the development of agentic frameworks. These systems are powered by generative models such as Large Language Model (LLM), that can autonomously orchestrate multiple tools within a unified environment, including data retrieval, geospatial analysis, and the application of machine learning and deep learning predictive models.
Retrieval-Augmented Generation (RAG) pipelines
In parallel, we develop Retrieval-Augmented Generation (RAG) pipelines, sequences of steps in which the system retrieves reliable, scientifically validated information from verified knowledge bases of documentation, which is then used by the AI to generate analyses, reports, and responses that are scientifically accurate, and coherent with the results provided by the other orchestrated tools.
MEEO application example
SMART project
SMART (Sensibilization and Mobilization for an Active Green Transition), an ESA Space Solutions Kick-start activity led by MEEO, explores how Large Language Models (LLMs) can turn Earth Observation and climate data into engaging, accessible content for a wide range of audiences, from expert communities to the general public and younger generations.
The core challenge it addresses is shared across research institutions, media, and public administrations: communicating complex climate-related information and actions in a way that is clear, relevant, and actionable.
SMART is built around two complementary LLM-powered workflows. In the first, a semantic interface lets users query Earth Observation and climate data in natural language: drawing on satellite imagery, climate models, reanalysis outputs, and socio-economic data held in MEEO’s geospatial data management platform, an LLM generates tailored data visualisations on demand. In the second, a dedicated LLM produces communication and report material by combining these results with verified knowledge sources, returning clear, contextualised outputs to the user. Across both, content is personalised to each user’s interests, context, and level of knowledge, with real-time relevance that links current events to long-term climate trends. The activity completed its technical and economic feasibility and defined an implementation roadmap towards a proof of concept, service customisation, and demonstration.