From predicting extreme weather events to ocean monitoring and biodiversity mapping, discover how research is revolutionising our understanding and protection of the environment.
At the heart of this innovation: hybrid AI, which reconciles physical models and ML for reliable, explainable, and operational predictions.
Meteorology and Climate – Hybrid AI Facing Extreme Events
The New Paradigm of Weather Forecasting
Historically, weather forecasts relied exclusively on complex physical models (NWP, Numerical Weather Prediction). These models simulate the laws of thermodynamics and fluid mechanics to predict atmospheric evolution. These models, the result of decades of research in atmospheric physics, have long constituted the gold standard of weather forecasting.
However, recent years have seen a radical paradigm shift, driven by new actors from the tech industry (Google, Huawei, NVIDIA, Microsoft). They have successfully developed approaches based purely on Machine Learning, exploiting the power of deep learning and vast archives of historical weather data.
These “data-driven” models often prove more accurate for general forecasts. Most importantly, they require much less computing power at prediction time (inference). They represent a significant advance in the operational efficiency of weather forecasting.
However, despite their impressive performance, these ML models present two major challenges that limit their complete adoption in critical operational systems:
1. Uncertainty Quantification
It is more challenging to evaluate and quantify the uncertainty associated with their predictions. Yet, this quantification is essential for controlled risk management, particularly in domains where forecasts inform critical decisions that impact public safety.
2. Extreme Events
They struggle to anticipate extreme events (hurricanes, flash floods, heat waves). By nature rare, these phenomena are underrepresented in training data, making them difficult to grasp for a purely statistical approach. Yet it is precisely these events that have the most serious human and material consequences.
We thus understand that ML models, while promising and performing well on standard cases, still struggle to fully address all the challenges of operational weather prediction.
| Criterion | Physical Models (NWP) | AI Models (Data-Driven) |
| Computing power (inference) | Very high | Low |
| General accuracy | High | Often superior |
| Extreme events | More reliable (physical laws) | Weak (rare data) |
| Uncertainty quantification | Well mastered | Complex to evaluate |
| Explainability | High (physical meaning) | Low (“black box”) |
The Solution: Hybrid AI
Comparing these two approaches, a natural question arises: why not combine them? This is precisely what so-called hybrid AI approaches seek to do. They represent today’s frontier of innovation in weather forecasting.
Hybrid AI? We already discussed it here: Future Intelligence
For example, we see physical models where certain components are replaced or enriched by AI modules, allowing calculations to be accelerated without sacrificing physical consistency. Other approaches use ML algorithms to increase the precision and resolution of traditional simulations, particularly in areas where physical measurements are missing or unreliable.
It is in this context that the EXPLEARTH chair (EXPLainable and physics-informed AI for Regional weaTHer prediction) fits. This promising project embodies this hybrid vision through two complementary axes:
Axis 1 – Development of Hybrid ML Models
Develop and evaluate innovative ML approaches, including:
- An ML version of the Arome model (Arome-AI), the French reference for medium-scale forecasting
- Spatial resolution improvement (downscaling) of weather predictions by physical models, allowing progression from regional forecasting to fine local forecasting
Axis 2 – Explainability and Analysis
Focus on model analysis to ensure its robustness and make it more understandable (explainability). The ambitious goal is to understand what ML models have actually learned, and then improve the physical models themselves. This “model inversion” approach creates a virtuous circle of mutual improvement.
Space Weather: A Case Study for Hybrid AI
This hybrid approach finds an echo in the field of solar weather, also called space weather. This sector, on which entities like ONERA (AI & DPHY Space Weather and Space Climate department) work, faces specific challenges. These challenges make the purely physical approach extremely complex and the purely data-driven approach insufficient.
Three major problems arise in this domain:
The physics of solar phenomena is not yet fully understood, limiting the ability of current models to faithfully capture it. The processes of solar flare generation and solar wind propagation involve complex magneto-hydrodynamic phenomena that are still partially mysterious.
Existing physical simulations are extremely computationally intensive and time-consuming, making their use impractical for real-time alert systems. A complete simulation can take hours, while the event propagates in just a few tens of minutes.
Observational data is very rare, fragmentary, and multimodal, meaning of very different natures: solar images in different wavelengths, in-situ particle measurements, radio signals, magnetic measurements. This heterogeneity makes it difficult to optimally exploit the available information.
The objective is to use Machine Learning in a targeted and intelligent way:
On one hand, to intervene where physical models fail or lack precision, by learning directly from patterns in observed data.
On the other hand, to develop rapid and operational alert systems. In many cases, physical simulations are indeed too slow to provide an exploitable alert in a useful time before a potentially dangerous solar event.
This application of hybrid AI to space weather demonstrates how the judicious combination of physical and data-driven approaches can overcome limitations that would be insurmountable with a single approach.
Oceans and Climate – High-Resolution Global Monitoring
A New Era of Ocean Observation
The application of AI to ocean systems deepens our understanding of climate and our ability to anticipate its evolution. At the heart of this change are next-generation satellite missions, particularly SWOT (Surface Water and Ocean Topography) and its KaRIn instrument (Ka-band Radar Interferometer).
Recent missions have provided two-dimensional maps of sea surface height at higher resolution and with more precise spatial coverage. They are capable of seeing structures at scales that could only be imagined with previous altimeters.
These new measurements open two doors at once:
- A finer observation capacity of oceanic phenomena
- A possibility to rethink how we integrate this massive and complex data into already costly numerical models
From Traditional to Hybrid Approaches
Physical Models and Assimilation: The Physical Approach
Ocean numerical systems rely on the fundamental laws of fluid dynamics and sophisticated numerical models that have been developed over several decades. Data assimilation plays the role of super-integrator of observations. It continuously recalibrates a model’s trajectory based on available measurements and has literally transformed ocean forecasting and historical reconstruction. These systems are now operational, robust, and constitute the backbone of global oceanographic services.
However, important practical limitations must be taken into account. These codes are computationally intensive, and although they perform, they are not perfect. They may misposition certain structures, exhibit systematic biases in specific regions, or lack the resolution to capture fine phenomena.
Machine Learning Models: The Predictive Approach
At the other end of the spectrum, neural networks applied to satellite products excel at concrete and targeted tasks. For example: image denoising, spatial and temporal interpolation, multi-sensor fusion, and local reanalysis of missing fields. They are much lighter and faster than physical models and can directly capitalise on observations and products from new satellites.
However, their dependence on the quality and coverage of training data prevents them from representing all complex oceanic processes and limits their extrapolation capacity outside their distribution. A network trained on normal conditions can fail completely when facing an exceptional event.
Hybrid Oceanographic AI: Architecture and Applications
This is where the hybrid approach, notably developed by CLS in collaboration with IMT-Atlantique, comes in. The ocean reconstruction they develop combines physical models and deep learning to map in real-time temperature, salinity, and currents at a global scale. These models facilitate the detection of marine heatwaves (increasingly frequent events associated with climate warming). They allow for improved climate predictions.
In use cases currently studied by CLS and IMT-Atlantique:
Error Correctors: A neural network learns the systematic residual bias of the physical model and applies it as a correction on the fly during simulation. Particularly useful for locally recentering oceanic structures that may be “offset.”
Ultra-Fast ML Substitutes: Replace particularly computationally expensive components (such as certain complex physical parameterisations) with an ML substitute that reproduces their behaviour.
Multi-Observation Fusions: CNN or transformer architectures with attention mechanisms to intelligently merge new high-resolution altimetry data with other sources. For example: satellite sea surface temperature, in-situ temperature and salinity profiles (Argo floats), current measurements, etc.
Agroforestry and Biodiversity – Mapping Ecological Oases
Revealing the Invisible: Agro-Ecological Infrastructures
The application of AI to the characterisation of agroforestry landscapes marks an evolution in our understanding and management of rural territories. The automated mapping of agro-ecological infrastructures: hedgerows, groves, permanent grasslands now allows identifying and inventorying with unprecedented precision these crucial but discreet elements for ecological connectivity and biodiversity preservation.
These small landscape structures have long been neglected in traditional conservation approaches, which focused on large forests and protected areas. Today, these small structures reveal their capital importance as true oases of biodiversity preservation in agricultural environments. Hedgerows, in particular, constitute essential ecological corridors. They are refuge zones for many species. They play a key role in local climate regulation, soil protection, and water quality.
Thanks to advances in satellite remote sensing (increasingly fine spatial resolution) and algorithms (notably semantic segmentation architectures), these zones can now be identified, mapped, and their evolution tracked systematically and reproducibly over vast territories.
Going Back in Time to Inform the Future
Beyond this mapping of the present state, innovations in artificial intelligence now enable the reconstruction of the historical evolution of agroforestry landscapes by analysing archival images, revealing the dynamics of deforestation, agricultural intensification, and habitat fragmentation over several decades. This ability to “go back in time” proves crucial for informing current and future development decisions.
The SegCycleGAN architecture developed by the DYNAFOR team (Dynamics and Ecology of Agroforestry Landscapes) perfectly illustrates this innovative capability. This deep learning architecture exploits unsupervised learning techniques (CycleGAN) combined with semantic segmentation to analyse unpaired historical archival images. That is, without needing to have manually annotated historical maps for each period.
The model learns to translate old images (often black and white, with variable resolution, with different acquisition conditions) into land use maps comparable to current maps. This capability allows reconstructing landscape evolution, revealing:
- Areas where hedgerows massively disappeared during agricultural consolidations
- Dynamics of landscape closure (scrub encroachment) or conversely opening (clearing)
- Trajectories of agricultural intensification or extensification
- Impacts of successive public policies on landscape structures
This fine understanding of territorial ecological trajectories allows planners and decision-makers to grasp long-term landscape transformation processes and anticipate their consequences on biodiversity and ecosystem services (pollination, pest regulation, water quality, carbon storage).
Mapping Old Forests, Biodiversity Sanctuaries
Parallel to this historical approach to agricultural landscapes, the automated mapping of old forest maturity poses a significant challenge for forest biodiversity conservation. These mature or senescent forest ecosystems now account for less than 2% of the European forest area. However, they constitute absolutely critical biodiversity reservoirs.
Old forests are characterised by:
- A complex vertical structure with multiple vegetation layers
- The presence of dead wood on the ground and standing, is essential habitat for thousands of species
- Old large-diameter trees, irreplaceable microhabitats
- A diversity of species and microhabitats (cavities, peeling bark, fungi)
- Little-disturbed natural ecological processes
Automated identification of these forests via LiDAR data (which measures vegetation’s 3D structure), multispectral satellite images, and sophisticated classification algorithms allows for systematic mapping. These approaches guide conservation strategies by precisely identifying:
- Forest areas with very high ecological value require strict protection
- Maturing forests that could be left to free evolution
- Ecological restoration opportunities in degraded areas
- Connectivity networks between old forest islands
New Technological Approaches
Beyond these concrete use cases, three innovations are redefining the contours of AI for the environment.
Physics-Informed Learning: Reconciling Physical Knowledge and AI
Physics-informed learning represents one of the most promising innovations in environmental AI. This hybrid approach, illustrated by Physics-Informed Neural Networks (PINNs) developed at ONERA and the work of the EXPLEARTH chair, directly integrates physical constraints into the architecture and loss function of neural networks.
The fundamental principle: Rather than letting the network freely learn statistical patterns, we impose compliance with certain differential equations, conservation laws, or symmetry principles from the physics of the problem. Concretely:
Hard Constraints: The architecture automatically satisfies constraints (e.g., mass conservation)
Soft Constraints: The loss function penalises violations of physical laws
Hybrid Components: Some parts are explicit physical modules, others are neural networks
The decisive advantages:
- AI’s learning capacity to capture complex and non-linear patterns
- The theoretical robustness of physical models to ensure scientific consistency
- Explainability: unlike “black boxes,” these models are based on verifiable natural laws, crucial for scientific validation and operational acceptance
Physics-informed learning particularly excels in managing uncertainties and extreme events. Physical models maintain their superiority during rare events by relying on fundamental principles valid even in never-observed conditions, while pure AI can fail on out-of-distribution events. Hybridisation allows intelligent role distribution according to contexts.
Foundation Models and Multi-Modal Fusion: Generalisation and Orchestration
The Revolution of Foundation Models for Earth Observation
Foundation models (alphaearth, terramind) represent a paradigmatic rupture. Instead of training a specialised model per task, generic models capable of solving varied tasks are developed. Once the foundation model is obtained (at the cost of massive initial training), it becomes significantly easier to develop new capabilities by leveraging the acquired general knowledge.
The SATLAS model, used by CLS for detecting woody elements, illustrates this spectacular evolution. From traditional ML requiring tens of thousands of annotations, we move to only 10,000 labelled thumbnails for fine adaptation. This drastic reduction in annotated data needs to solve a critical problem: the scarcity of expert data in many geoscientific domains.
The Necessary Specialisation
Unlike text or natural images, geospatial data requires architectures adapted to multiple modalities: visual (RGB), hyperspectral (hundreds of bands), radar (SAR), LiDAR (3D clouds), each with unique physical specificities.
DYNAFOR evokes the perspective of multi-modal foundation models. They allow generalisation to different regions, scales, and sensors, potentially revolutionising landscape ecology through global comparative analyses with unified methodologies.
Multi-Modal Fusion: Intelligent Sensor Orchestration
Each modality captures specific complementary aspects: visible and near-infrared for vegetation, hyperspectral for chemical composition, radar for structure, LiDAR for 3D topography. No single modality can capture the entirety of information.
The MIPANet architecture (DYNAFOR) uses cross-attention mechanisms to orchestrate this complementarity. These Transformer architectures allow:
- Dynamically weighting the importance of each modality according to context
- Automatically identifying which modalities are most informative
- Elegantly handling missing data (clouds for optics)
- Extracting subtle inter-modal correlation patterns
Collaboration Between Physical Models and ML Models
A new approach seeks to go further than explainability, in the classic sense, by transforming the relationship between physical modelling and Machine Learning:
- Physical models provide constraints and scientific structure to AI
- AI reveals patterns in data and identifies systematic errors in physical models
- These insights are reinjected to improve physical models themselves
- A virtuous circle of mutual improvement is created
Concrete Applications at ONERA: Work on atmospheric signal inversion demonstrates this capability. By systematically analysing gaps between physical predictions and observations, then using AI to identify patterns, researchers can discover neglected or poorly parameterised physical processes and refine models.
Inference on Incomplete Data: A notable feature is the capacity for robust inference on incomplete data. AI models excel at producing coherent results where physical models require complete datasets. This capability allows filling observational gaps by relying on learned patterns and physical constraints.
Operational Performance: The EXPLEARTH chair demonstrates, through its meteorological emulators, the quasi-real-time inference capacity of ML models, which is several orders of magnitude faster than traditional physical models. Where a physical forecast requires hours on a supercomputer, the ML emulator produces an equivalent forecast in a few seconds. These approaches facilitate the use of critical real-time applications.
Conclusion
The Toulouse ecosystem demonstrates the complementarity between AI and traditional scientific approaches. The hybridisation of Machine Learning models and physical models enables the identification of complex patterns while maintaining consistency with physical laws and quantifying uncertainties.
The work conducted identifies three major research axes:
Uncertainty Management: Combine the robustness of physical models and the efficiency of ML to develop reliable systems under varied conditions
Explainability: Ensure transparency and interpretability of hybrid models, particularly for critical applications
Climate Change Adaptation: Maintaining model performance in the face of unprecedented climate conditions, outside the distribution of training data.
AI4ENV is led by ANITI. This initiative is part of ANITI’s (Artificial and Natural Intelligence Toulouse Institute) strategy, a centre of excellence in trustworthy AI led by Serge Gratton, which brings together 300 team members, 19 research chairs, and 8 integrative programs with a budget of € 90 million.
ANITI’s originality lies in its approach to trustworthy AI for critical applications, combining fundamental research (learning architectures, multimodal learning) with essential attributes (explainability, robustness, frugality) in three major integrative domains: transportation, Industry 4.0, and the environment.