Teaching AI to See Rain: How Satellite Data Is Transforming Precipitation Forecasting

The most reliable precipitation forecasts in the world still struggle with a blind spot: they can't see rain the way satellites can. Conventional weather models rely on reanalysis datasets—best guesses about the atmosphere stitched together from sparse observations—which tend to smooth out or misrepresent the localized bursts of rainfall that matter most for daily life. Now, a team of researchers has shown that AI weather models trained on actual satellite observations of rainfall produce forecasts that beat these traditional baselines, and that feeding raw satellite data directly into the model during forecasting provides an additional boost at short lead times. The finding suggests that the path to better rain prediction runs through space.

The Science

When you check your phone for tomorrow's chance of rain, you're relying on one of two things: a numerical weather prediction (NWP) model that solves physics equations describing atmospheric motion, or an AI model that has learned from decades of atmospheric data how weather patterns tend to evolve. Both approaches have strengths, but both share a fundamental weakness when it comes to precipitation: they struggle to see what's actually happening with water in the atmosphere.

The challenge is that rain is messy. It forms through microphysical processes—water vapor condensing into droplets, droplets growing heavy enough to fall—that happen at scales smaller than even the most powerful supercomputers can fully resolve. A thunderstorm that dumps two inches of rain in thirty minutes might span just a few kilometers and last less than an hour. Capturing that requires either extraordinary computational resolution or direct observations of rainfall itself.

Conventional NWP models try to simulate these processes using parameterizations—mathematical approximations of what physics says should happen—rather than observing it directly. AI weather prediction (AIWP) models like Pangu-Weather, FourCastNet, and GenCast have made enormous strides in forecasting temperature, pressure, and wind patterns, often matching or exceeding conventional models. But precipitation has remained a secondary concern. Most AI weather models learn to forecast from reanalysis datasets, which themselves contain considerable uncertainty, especially for precipitation.

The research team, based at institutions including Colorado State University and IBM Research, took a different approach. Rather than relying solely on reanalysis inputs to infer precipitation, they asked two questions: What if we trained the AI on satellite observations of actual rainfall? And what if we fed the model real satellite data at prediction time, not just during training?

Their system, called Prithvi-Precip, builds on Prithvi-WxC, a transformer-based AI foundation model pre-trained on MERRA-2 reanalysis data covering two decades of atmospheric conditions. Transformers—the same architecture behind large language models—excel at finding patterns across long sequences of data, making them well-suited for tracking how weather systems develop over time.

The Prithvi-WxC model operates on a global grid with roughly 0.67° resolution in the east-west direction and 0.5° in the north-south direction—comparable to about 75 kilometers at the equator. The model ingests 20 surface-level variables (temperature, pressure, wind components, humidity) and 10 vertically-resolved variables at 14 different altitude levels, giving it a detailed picture of the atmosphere's three-dimensional structure.

To predict precipitation, the researchers attached a specialized "precipitation head" to the model—an additional neural network component that takes the model's understanding of atmospheric state and translates it into a rainfall forecast. The architecture is illustrated in

.

The real innovation lies in how the model handles satellite observations. Satellite data comes from wildly different sensors: geostationary satellites like GOES sit 35,000 kilometers above the equator and take continuous images in visible and infrared wavelengths, providing excellent temporal coverage but limited information about precipitation deep inside clouds. Polar-orbiting satellites like those in NASA's GPM constellation pass overhead at much lower altitudes, carrying microwave sensors that can peer through clouds to sense precipitation directly, but cover only narrow swaths at any given moment. The GPM Microwave Imager (GMI), for instance, has 13 channels spanning frequencies from 10.65 GHz to 183 GHz, each sensitive to different aspects of precipitation and atmospheric water vapor.

shows examples of the data these different sensors provide. The challenge is that an AI model typically expects a fixed set of input features, but satellite observations arrive in variable quantities and formats depending on which satellites happen to be overhead at any given time.

The researchers solved this with a sensor-agnostic observation encoder. Each satellite channel—whether it's a GMI channel at 23.8 GHz or a GOES infrared window at 11 micrometers—is treated as an independent "observation layer" with its own metadata (wavelength, polarization, observation time). An encoder maps these variable-length observation sequences into a fixed-dimensional representation that can be combined with the model's internal atmospheric state. This approach, inspired by the Perceiver architecture, uses cross-attention to summarize all available observations for each location into a fixed-size vector.

shows the observation encoder architecture in detail. Observations are processed at the tile level—the internal grouping used by the Prithvi-WxC model—and then combined across time steps before being merged with the atmospheric state. The final architecture, shown in

, concatenates the encoded observations with the atmospheric state, passes them through a small neural network that computes a correction, and adds this correction back to the atmospheric representation. A separate precipitation head then produces the rainfall forecast.

What They Found

The researchers systematically evaluated key design choices, and the results upended some assumptions about how best to train AI precipitation models.

The first comparison was between two forecasting strategies: autoregressive rollout and continuous forecasting. In continuous forecasting, the model is trained to predict precipitation directly at any requested lead time—you ask for the 72-hour forecast, and it produces one. In autoregressive forecasting, the model predicts six hours ahead, then takes that prediction as input to predict twelve hours ahead, and so on, rolling forward like a domino chain.

Autoregressive approaches had seemed computationally inefficient and potentially prone to accumulating errors. The conventional wisdom in AI weather prediction favored continuous forecasting because these models also predict the full atmospheric state, which provides a physical constraint that helps keep errors from growing. Precipitation, however, is not part of that atmospheric state in Prithvi-Precip—it's predicted separately.

The results were striking. As shown in

, autoregressive forecasts substantially outperformed continuous forecasts across every metric: lower normalized root mean squared error (NRMSE), higher correlation coefficients, and better continuously ranked probability scores (CRPS). The continuous model showed significant positive bias at short lead times, predicting more rain than actually occurred, while the autoregressive model remained nearly unbiased out to 36 hours.

This finding has practical implications for how AI precipitation systems should be trained. Even without the physical consistency constraints that help other weather variables, recursively predicting forward in time produced more accurate results than predicting all lead times at once.

The second major comparison examined how the choice of training target affects forecast skill. The researchers trained identical models on two different precipitation datasets: MERRA-2 reanalysis precipitation and IMERG V07 satellite-derived precipitation estimates. IMERG (Integrated Multi-SSatellite Retrievals for the Global Precipitation Measurement) combines observations from multiple satellite sensors—including microwave sounders and radar—to produce what are generally considered the most accurate global precipitation estimates available, calibrated against surface rain gauges where possible.

When evaluated against the data they were trained on, both models performed similarly well. But the critical test came when they were evaluated against independent observations: gauge-corrected radar precipitation estimates from the NOAA Multi-Radar Multi-Sensor (MRMS) system over the contiguous United States.

shows that models trained on IMERG consistently outperformed those trained on MERRA-2 when evaluated against independent radar data. The improvement was substantial across all lead times tested, from 6 to 96 hours. This demonstrates that the quality of the training target matters enormously for precipitation—IMERG's satellite-derived rainfall estimates capture real precipitation patterns more accurately than MERRA-2's reanalysis-derived values, even when the evaluation data is completely independent of both.

This finding has important implications for how AI weather models should be developed. The convention in the field has been to train on reanalysis because it's consistent with the atmospheric state variables used for other weather parameters. The researchers' work suggests this is a suboptimal compromise for precipitation specifically.

The third set of experiments tested the impact of directly ingesting satellite observations into the forecasting model. Here the results were more nuanced, but still meaningful.

compares forecasts from models with and without satellite observation ingestion. At short lead times (6-24 hours), the satellite-fed models showed clear improvements in skill scores, particularly in tropical and subtropical regions where satellite observations are most informative—these are areas where conventional data assimilation struggles due to sparse conventional observations and complex convection.

However, the benefits of satellite ingestion diminished at longer lead times. By 72-96 hours, the advantage was marginal. This makes physical sense: satellite observations provide information about the current state of the atmosphere, which matters most for near-term predictions. As forecasts extend further into the future, the chaotic evolution of the atmosphere dominates, and initial conditions matter less.

To create a model that can operate even when satellite observations are incomplete or missing, the researchers trained with random dropout—each observation layer had a 20% chance of being excluded, and a random 20% chance that the entire satellite input would be dropped. This made the model robust to gaps in observation coverage.

The final validation came from comparison with operational baselines. Prithvi-Precip forecasts were compared against the Goddard Earth Observing System (GEOS) forward processing system, which represents the most directly comparable conventional forecast since it uses the same underlying model and data assimilation system as the MERRA-2 reanalysis used to train Prithvi-Precip. Prithvi-Precip substantially outperformed GEOS across the full 96-hour forecast range.

Perhaps more impressively, the researchers also compared against the ECMWF's AI Forecasting System (AIFS), one of the strongest AI weather models currently in existence, which is initialized from the analysis produced by the IFS data assimilation system—arguably the world's most sophisticated weather analysis. Prithvi-Precip remained competitive, particularly at short lead times where the satellite ingestion showed its strongest benefits.

Why This Changes Things

Precipitation forecasting matters more than most people realize until they're caught without an umbrella—or worse, until a flash flood sweeps through their neighborhood. Agriculture depends on rainfall forecasts to decide when to plant, irrigate, or harvest. Energy traders anticipate hydroelectric output based on expected precipitation across watersheds. Emergency managers evacuate communities based on forecasts of heavy rainfall that could trigger landslides or compound flooding from swollen rivers.

The economic value of improved precipitation forecasting is difficult to calculate precisely, but studies have suggested that weather-related economic losses in the United States alone exceed $100 billion annually, with a significant portion attributable to excessive or poorly-timed precipitation. Even marginal improvements in forecast skill—being right 60% of the time instead of 55%, or predicting the right amount of rain 70% instead of 65%—could translate into billions of dollars in avoided losses.

The implications of this research extend beyond immediate forecast improvements. The field of AI weather prediction has been progressing rapidly, but most advances have focused on temperature, pressure, and wind—variables that are well-represented in reanalysis datasets and relatively straightforward to verify. Precipitation has lagged behind not because it's unimportant, but because the training and evaluation are genuinely harder.

This work points toward a broader principle: AI weather models should be trained on data that's actually representative of what we're trying to predict, not just what's convenient. Reanalysis was developed as a consistent historical record for training and verification, but it's essentially a model output itself, filtered through the assumptions and biases of the underlying forecast system. Satellite precipitation estimates, by contrast, are direct observations of what actually falls from the sky.

The observation encoder architecture also represents a significant advance in making AI models work with real-world data. Conventional data assimilation—the process of combining observations with forecasts to produce an initial condition—requires extensive hand-engineering to specify how different observation types should influence the analysis. The sensor-agnostic approach used here learns this mapping automatically, potentially making it easier to incorporate new satellite instruments as they come online.

There are implications for global health as well. Precipitation patterns influence the spread of vector-borne diseases like malaria and dengue, which are driven by standing water where mosquitoes breed. More accurate rainfall forecasts could provide early warning of conditions likely to produce disease outbreaks.

For the growing number of countries developing national AI weather capabilities, the work demonstrates that starting with foundation models trained on reanalysis is a reasonable first step, but that specialized improvements for specific variables like precipitation are achievable. The modular architecture—foundation model plus observation encoder plus precipitation head—suggests a path toward customized applications that inherit the general atmospheric understanding of a foundation model while optimizing for specific forecasting needs.

What's Next

The researchers are candid about limitations and open questions. The model was trained on data from 2000-2020 and validated on 2022 data, with evaluation against 2025 forecasts—a substantial temporal gap that tests but doesn't fully prove generalization to changing climate conditions. The evaluation focused on the contiguous United States and Brazil, leaving open questions about performance in other climate regimes, particularly the high latitudes and maritime regions.

The model predicts six-hour accumulated precipitation, which captures synoptic-scale weather systems well but may miss brief, intense rainfall events that last less than a few hours. The spatial resolution of roughly 0.67° by 0.5° is coarser than the resolution of the underlying satellite observations and coarser than what would be needed to capture orographic precipitation or other localized effects.

Perhaps most significantly, the benefits of satellite observation ingestion were strongest at short lead times but diminished substantially beyond 48 hours. This suggests that satellite data provides valuable information about current conditions but doesn't fundamentally change the chaotic nature of atmospheric evolution at longer ranges. Breakthroughs in extended-range precipitation forecasting may require advances beyond just better initial conditions.

The researchers note that their approach of training on satellite-derived precipitation targets rather than reanalysis may introduce its own biases, since IMERG and other satellite precipitation products have known limitations—they can struggle with orographic precipitation, light rain in high latitudes, and snowfall. No precipitation estimate is perfect, and choosing which imperfect estimate to train on involves tradeoffs.

Several avenues for future work emerge naturally from this research. Ensemble forecasting—running multiple forecasts with small perturbations to estimate forecast uncertainty—has proven valuable for conventional weather prediction but remains underexplored for AI precipitation systems. The current work focused on global precipitation, but specialized models for extreme events, seasonal prediction, or specific regions might benefit from the same principles. And the observation encoder could potentially be extended to incorporate non-satellite observations like surface weather stations or aircraft reports.

For now, Prithvi-Precip represents a proof of concept that AI precipitation forecasting can be substantially improved by taking satellite observations seriously—not just as verification data, but as training targets and as direct inputs to the forecasting model. Whether these gains hold up in operational deployment, across longer time periods, and in the diverse climate regimes where rain matters most will determine whether this approach becomes standard practice or remains an interesting research result.

The broader lesson may be that AI weather prediction is still young enough that assumptions about how to build these systems are being tested and revised. The convention of training on reanalysis, predicting continuously rather than autoregressively, and focusing on atmospheric state variables rather than precipitation directly—all of these reflected reasonable prior beliefs that are now being challenged by systematic experimentation. As the field matures, the best practices for each variable and application will become clearer, and precipitation—the most consequential weather event for most people's daily lives—deserves its place as a primary focus rather than an afterthought.