The problem
A radar map goes blank where no radar reaches, and a clear sky looks the same. Satellites see everywhere, so we designed a model to fill that gap from them. Its output is satellite-derived: an estimate of radar reflectivity, not a radar measurement, and despite the chapter’s title, not a Doppler velocity.
Every map would carry an Observed or Satellite label, so whoever reads it knows which one they are looking at. The planned global product updates every 10 minutes at roughly 4 km. Everything below is a design; none of it is a measured result. Source §4.
The data
An infrared image measures radiation from the cloud top. Brightness temperature puts that radiance on a temperature scale. It is useful information and no direct measurement of precipitation under the cloud. The same apparent cloud top sits above very different conditions lower down.
The proposed targets come from radar composites where radar exists: MRMS in CONUS, OPERA in Europe, AURA in Australia, IDEAM in Colombia, REDEMET in Brazil, and Taiwan CWA. A paired training example has to align the satellite input and the radar observation in space and time. Source §4.
The model
The design uses three frames from t−20 through t, visible imagery by day, and solar-zenith and terrain information alongside. GFS fields arrive on a 0.25° grid and cover atmospheric water, freezing level, instability and inhibition, wind shear, column humidity, and model precipitation rate. Each is a separate physical clue to what lies under the cloud top. Source §4.
A U-Net joins local image features to broader context and returns an output at each pixel. The planned size is roughly 10–25 million parameters, with an encoder shaped like the nowcasting model’s. It predicts class scores instead of one unconstrained number. The design also has separate day and night versions and forecast heads for 30 and 60 minutes ahead. Source §4.
An ambiguous input should be able to say that rain is uncertain. Averaging incompatible target values with squared error gives a “gray soup”: a middling value that matches no clear sky and no strong echo. Assigning probability to classes states the no-rain possibility outright and still lets probability spread across rain strengths. Source §4.
Where this comes from
The architecture is not novel. Infrared-to-rain learning goes back to Hsu et al.’s PERSIANN (1997). Much closer is Hilburn, Ebert-Uphoff, and Miller’s GREMLIN (2021): satellite temperatures plus lightning mapped to MRMS reflectivity. Lee and Hilburn (2024) document its winter cold-surface failure. What we would add is training data from several countries’ radar networks, the weather-model environment fields, the Observed/Satellite labeling, and the forecast heads. Until the model is trained, none of those additions has a measured value. Source §6.
The loss and why that loss
For the CSI event score, $\mathrm{CSI}=\mathrm{hits}/(\mathrm{hits}+\mathrm{misses}+\mathrm{false\ alarms})$, from zero (no overlap) to one (perfect overlap). The small synthetic model in trainkit scores 0.999 against 0.203 for its climatology baseline: almost every toy event pixel agrees, against about one fifth for the baseline. That toy says nothing about how the satellite model would score. CSI ignores how severe a miss is, so the design also checks the stronger reflectivity threshold. Source §6.
$$L=-w_y\ln p(y\mid\mathbf{x})+L_{\text{CSI-aligned}}$$In words: penalize low probability on the observed class, multiply by that class’s weight, and add the planned event-score penalty.
Here x is the input feature collection, y is the observed class, p is its predicted probability, and w is the class weight. Cross-entropy gives a gradient even when the final display would round several different distributions to a similar shade. The additional CSI-aligned term focuses attention on the event thresholds used for evaluation. Source §4.
As in the loss chapter, the choice of objective encodes a priority. It does not prove the priority was achieved. We still need held-out predictions to find whether weighting increases useful detections or merely increases false alarms.
Training as it ran
We designed this model but did not train it, so there is no epoch curve, learning rate, batch size, training time, or held-out score. We are still assembling the training data. Source §4.
The planned test scores CSI in held-out regions at 20 and 35 dBZ against climatology, then repeats the comparison without the GFS environment block. Withholding regions asks whether the mapping travels past the places that supplied training radar. Removing an input group asks what that group contributes under a controlled comparison. Source §4.
Outputs
The widgets below run the small synthetic model from trainkit, not the satellite model. Clicking a location on the infrared image picks a cell of that toy.
Move the selected cell and watch the probability bars. The toy has only a few broad classes; the display spreads them across the finer reflectivity bins the satellite model would use, to show what that chart looks like. The bars are not a learned distribution over those bins, and the infrared frame is there for orientation.
A wide distribution is no calibrated confidence statement on its own. A model can be confidently wrong outside its training conditions. Holding out whole regions matters because the places that need satellite-derived estimates can have different terrain, cloud regimes, or observing geometry from the radar-covered training sources.
What it is not
Climatology is the baseline: the usual frequency or pattern of the event, a reference that doesn’t lean on the current satellite image the way the model does. A useful learned product has to improve on it in the held-out region. Removing the environment inputs then asks whether the extra complexity is worth it. Source §4.
Where this shows up when you train
Write down the baseline and the test before you train, and keep a planned experiment apart from a finished result.
Before publishing a satellite-derived image, we would also have to establish what happens when an input goes missing. A daytime visible channel cannot quietly become a valid nighttime observation. That is why the design has separate day and night versions. Having them on paper is not the same as having trained and tested them, and the interface still has to keep the observation source and the model status legible to whoever reads the map.
Worked example
For an illustrative three-class probability vector (0.5, 0.3, 0.2), suppose the first class is observed and its weight is 1. Its cross-entropy term is $-1\cdot\ln(0.5)=\ln(2)\approx 0.693$. Assigning that class probability 1 instead gives $-\ln(1)=0$. The loss rewards probability on the observed class, not the tallest-looking bar. Source §7.
Run this yourself
trainkit rainmap samples/ir-rain.npz --output rainmap-output
The command writes run.json and model.json. Add --dry-run to preview the replacement targets. Point --output at a folder of your own. The run records the data shapes, split, objective, baseline, model score, and elapsed time; the model export holds the learned weights themselves.
Worked example
When we ran it, the toy scored a CSI of 0.999 for its rain and heavy-rain classes taken together, against 0.203 for the climatology baseline, and trained in about 7 seconds. Source §6.
The gain is $0.999-0.203\approx 0.80$. It scores the toy’s rain classes, not a real reflectivity threshold. We ran it once, so there is no interval on the gain. Source §6.
Repeat the run before you change the recipe. Then change one input or model choice and compare on the same held-out examples. A different random split moves a score even when the learning rule is untouched. Keeping the original baseline is what separates an improvement in the learner from an easier evaluation problem.
What you now know
- Satellite-derived reflectivity is an estimate, not an observed radar return or Doppler velocity.
- A distribution can express competing rain possibilities that one average hides.
- The architecture follows published work; only a held-out test can show whether it works here.
- The next evidence must come from held-out regions and an input ablation; the trainkit toy cannot supply it.
Where we’re headed
Return to Designing the Run to turn a proposed model into a test you can run.