Satellites See Clouds. This AI Sees Rain — Even the Extreme Kind

When a storm brews over rural France, farmers need to know not just that rain is coming, but how much — especially if it could trigger flash floods or wash away topsoil. But satellites don’t measure rainfall directly. They see cloud tops glowing in infrared light, cold and bright, and from that, scientists must infer what’s happening thousands of feet below, where raindrops form and fall. For decades, this has been a guessing game with high stakes. Now, a new study shows that the best way to turn satellite images into accurate rainfall maps may not be the most obvious one: sometimes, the AI that gets the individual pixels ‘wrong’ actually gets the storm right.

In a sweeping comparison of deep learning models, researchers from Météo-France and CNRS tested four state-of-the-art architectures — deterministic and generative — to estimate rainfall at high resolution using only infrared brightness temperatures from the Meteosat satellite. Their findings, based on 15 years of data across metropolitan France, reveal a fundamental trade-off: models that excel at pixel-level accuracy consistently underestimate extreme rainfall, while those that embrace uncertainty produce less precise images but far more realistic storms (Meignin et al., 2026).

This isn’t just a technical detail. It’s a paradigm shift in how we think about environmental AI. For applications like flood forecasting or climate risk modeling, getting the extremes right matters more than perfect pixel alignment. And that means the future of weather prediction may lie not in deterministic certainty, but in probabilistic imagination.

The Science

The challenge of estimating rainfall from space begins with physics. Geostationary satellites like Meteosat Second Generation (MSG) continuously monitor Earth’s atmosphere, capturing infrared (IR) radiation emitted by cloud tops. Colder clouds appear brighter in these images — a sign of high, thick storm systems. But cold cloud tops don’t always mean heavy rain. Some towering clouds produce little precipitation; others, like mesoscale convective systems, unleash torrents despite modest cloud-top temperatures.

This weak and nonlinear relationship has long limited quantitative precipitation estimation (QPE) from IR data. Traditional methods, such as the nowcasting satellite (NWC) algorithms, rely on empirical thresholds and lagged correlations. More recent approaches use machine learning to learn the mapping from IR to rain — but most treat it as a deterministic regression problem: given an IR image, predict the corresponding rainfall field.

Meignin et al. questioned this assumption. They hypothesized that because rainfall is intermittent and highly skewed — most pixels are dry, a few experience extreme events — a deterministic model might smooth over rare but critical outcomes. Instead, they proposed comparing deterministic models with generative ones, which can sample from a distribution of possible rainfall fields conditioned on the same IR input.

Their dataset spans 2008 to 2023, combining MSG’s 12-channel SEVIRI infrared observations with high-resolution radar mosaics from Météo-France, which serve as ground truth. The radar data are aggregated to a 3 km × 3 km grid at 15-minute intervals, co-registered with the satellite imagery. To handle the extreme class imbalance — over 90% of pixels record no rain — the team implemented stratified sampling, ensuring that training batches include a representative proportion of moderate and heavy rainfall events.

They evaluated four model families:

• Deterministic U-Net: A convolutional encoder-decoder with skip connections, widely used in image-to-image translation tasks.
• Transformer-based model: A vision transformer (ViT) adapted for spatial prediction, leveraging self-attention to capture long-range dependencies in cloud systems.
• Conditional Generative Adversarial Network (cGAN): A generator that produces rainfall fields, trained adversarially against a discriminator that learns to distinguish real from synthetic radar maps.
• Diffusion model: A score-based generative model that iteratively denoises a random field into a plausible rainfall pattern, conditioned on IR input.

All models were trained to predict rainfall rate (mm/h) from a sequence of three consecutive IR frames (t−30, t−15, t), enabling motion-aware nowcasting. Evaluation was performed on a held-out test set covering summer months (May–September), when convective rainfall dominates.

What They Found

The results revealed a stark divergence in model behavior. Deterministic models — both U-Net and the transformer — achieved the highest pixel-wise accuracy. They minimized mean squared error (MSE) and maximized structural similarity (SSIM), producing sharp, visually coherent rainfall maps that closely matched the radar observations in location and intensity for light to moderate rain.

But when it came to extreme events — defined as rainfall exceeding 20 mm/h — they consistently underpredicted. The U-Net, for example, captured only 62% of observed rainfall volume in the top 1% of pixels. The transformer did slightly better at 67%, but still fell short.

In contrast, the generative models captured the tail of the distribution far more faithfully. The cGAN recovered 89% of extreme rainfall volume, while the diffusion model reached 93% — a dramatic improvement (

Extreme Rainfall Volume Captured by Model Type

Label	Value
U-Net	62
Transformer	67
cGAN	89
Diffusion	93

Diffusion models recover nearly all extreme rainfall volume, while deterministic models miss over a third.

). This wasn’t due to overprediction: both models maintained reasonable false alarm ratios. Instead, they learned to ‘imagine’ plausible high-intensity configurations that deterministic models averaged out.

Extreme Rainfall Volume Captured by Model Type

Label	Value
U-Net	62
Transformer	67
cGAN	89
Diffusion	93

Diffusion models recover nearly all extreme rainfall volume, while deterministic models miss over a third.

More subtly, the spatial structure of rainfall differed. Deterministic models produced smoother, more diffuse precipitation fields — a known artifact of L2 loss minimization, which penalizes large errors heavily and favors conservative estimates. Generative models, by contrast, generated sharper gradients, isolated convective cores, and realistic spatial intermittency, closely resembling the fractal-like patterns seen in real storms (

).

This was particularly evident in case studies of mesoscale convective systems. When a large storm complex moved across central France in July 2021, the U-Net predicted widespread moderate rain (5–10 mm/h), missing the localized downpours that caused urban flooding in Clermont-Ferrand. The diffusion model, while less pixel-accurate, generated multiple plausible realizations, several of which included intense cells in the correct region — effectively providing probabilistic early warning.

To quantify this, the authors introduced a new metric: extreme event recall at 95% precision. At this threshold, the diffusion model detected 78% of observed extreme events, compared to just 41% for the U-Net. Even the transformer, with its global attention, only reached 49%.

Another key finding was calibration. Deterministic models were poorly calibrated: when they predicted 15 mm/h, the observed median was only 8 mm/h. Generative models, particularly the diffusion approach, showed better reliability curves, meaning their predicted intensities aligned more closely with reality when sampled across many events.

Why This Changes Things

These findings challenge a foundational assumption in environmental AI: that pixel-level accuracy should be the primary goal. In image restoration or medical imaging, minimizing MSE makes sense — a blurry tumor boundary could be dangerous. But in hydrology, a slightly misplaced but intense rainfall cell may matter more than a perfectly aligned drizzle.

Consider flood risk. A river basin’s response to rainfall is nonlinear: a 50 mm/h downpour on saturated soil can trigger a 100-year flood, while 20 mm/h over dry ground may cause no runoff at all. If AI systems systematically underestimate such extremes, they will underpredict flood probability — with real consequences for emergency planning and insurance.

This is not a hypothetical concern. In 2023, a slow-moving storm over eastern France dropped over 180 mm in 24 hours, causing deadly flash floods. Retrospective analysis using the U-Net model showed it would have predicted only 110 mm — a 40% underestimation. The diffusion model, by contrast, produced ensemble members exceeding 160 mm in multiple runs, flagging the event as high-risk.

Globally, over a billion people rely on satellite-based rainfall estimates because they lack ground radar. In sub-Saharan Africa, South Asia, and small island nations, these data inform drought response, crop insurance, and disaster preparedness. Yet most operational systems — including NASA’s IMERG and EUMETSAT’s NWC-SAF — use deterministic or blended approaches that likely suffer from the same bias.

The implications extend to climate science. Climate models project increases in extreme precipitation under global warming, but validating these trends requires long, consistent observational records. If satellite-derived rainfall datasets systematically dampen extremes, they may obscure the very signal we’re trying to detect.

Meignin et al.’s work suggests a path forward: embrace uncertainty. Generative models don’t claim to know the single true rainfall field. Instead, they acknowledge that multiple outcomes are possible given the same cloud-top observation — and they sample from that distribution. This aligns with how meteorologists think: a forecaster doesn’t say “it will rain 15 mm/h at 3:15 PM,” but “there’s a 60% chance of heavy thunderstorms this afternoon.”

The trade-off is computational. Generating 10 ensemble members from a diffusion model takes 15 seconds per patch, compared to 0.2 seconds for a U-Net. But for high-impact decision-making — evacuations, dam releases, crop spraying — that cost may be justified.

It also opens the door to hybrid systems. One could use a deterministic model for routine nowcasting and switch to generative sampling when conditions favor convection — high instability, strong moisture convergence — or when initial predictions exceed a threshold.

What’s Next

The study has limitations. It focuses on metropolitan France, a region with excellent radar validation data. How well these findings generalize to tropical or arid climates remains to be seen. Convective processes in West Africa or the Amazon may challenge even generative models.

The authors also note that their generative models were trained to match the statistical distribution of rainfall, not physical consistency. A sampled storm might have realistic intensity and structure but violate mass or energy conservation. Future work could integrate physics-informed losses or hybrid modeling, where generative networks are constrained by fluid dynamics.

Another frontier is temporal coherence. The current models predict single time steps. For nowcasting, one needs sequences. While the diffusion model can be unrolled autoregressively, this risks error accumulation. Video diffusion models or latent dynamics networks may offer better solutions.

Perhaps the most urgent next step is operational integration. Météo-France is already piloting a cGAN-based system for flash flood guidance. If successful, it could become part of the European Precipitation Index, used across 30 countries.

Meanwhile, the broader AI community is taking note. The paper’s open dataset and reproducible framework — including code for preprocessing, training, and evaluation — lower the barrier for others to build and compare models. That’s critical, because the best solution may not be any single architecture, but an ensemble that combines the precision of deterministic models with the variability of generative ones.

In the end, this study is about more than rain. It’s about how we design AI for high-stakes, uncertain domains. In medicine, finance, and climate, we often demand certainty from models that operate in inherently probabilistic worlds. Meignin et al. remind us that sometimes, the most accurate answer isn’t a number — it’s a distribution. And the most responsible AI isn’t the one that gets every pixel right, but the one that prepares us for the worst we might face.

$$\text{Precision}=\frac{\text{True Positives}}{\text{True Positives}+\text{False Positives}},\text{Recall}=\frac{\text{True Positives}}{\text{True Positives}+\text{False Negatives}}$$