The problem
A radar loop shows rain moving now. What goes on the map after the last observed frame? A nowcast is a short-range forecast. Reflectivity (dBZ) is the echo strength these fields show. We predict that quantity, and not the weather as a whole.
A lead time says how far ahead the prediction looks. The task asks for a field every 5 minutes out to 60 minutes. A forecast means something only against what a simpler method already does from the same observations. Source §3.
The data
Each input frame contains reflectivity, a valid-data mask, and horizontal motion components. optical flow estimates that motion from the image sequence. advection carries the existing echoes forward. The advected field is an explicit base prediction, not a hidden target. Source §3.
The model reads no infrared imagery. Its input is recent radar and motion, plus an advected base. Adding infrared is a separate, later experiment. Source §3.
Training uses 13,208 sequences and validation 441, with 34 whole days held out. No day appears in both splits, and no validation sequence starts within an hour of a training one. As in the split chapter, a lot of neighboring pixels is not a lot of independent weather situations. Source §3.
The model
The architecture joins local convolutional features to a carried recurrent state. A ConvGRU updates a grid of remembered features as frames arrive, so local spatial relationships survive while information accumulates over time. A U-Net gathers context at coarser scales, then folds it back into finer features while decoding the output.
FiLM scales and shifts features according to the requested lead. The lead embedding here uses 8 sinusoidal bands and width 128. The full model has 27,223,297 parameters, and the recurrent stacks hold 77 percent of them. The output is a residual added to the advected base, so the network edits that forecast instead of starting from an empty field. Source §3.
Where this comes from
The architecture is not novel: our convolutional recurrence follows Shi et al.’s ConvLSTM (2015) and TrajGRU (2017). We compare with advection and S-PROG in Pulkkinen et al.’s pySTEPS (2019). DGMR (Ravuri et al., 2021), the MetNet family, and NowcastNet (Zhang et al., 2023) are state-of-the-art references we do not claim to match. MetNet-1 (2020) and MetNet-3 (2023) are arXiv preprints; MetNet-2 (2022) is peer-reviewed. What we added is the training data, the regional fine-tunes, and the evaluation on held-out whole days. Source §6.
The loss and why that loss
critical success index (CSI) measures event overlap using the boxes above. Correctly dry pixels do not enter the denominator.
$$\operatorname{CSI}=\frac{\text{hits}}{\text{hits}+\text{misses}+\text{false alarms}}$$In words: divide correct event predictions by all pixels where either the prediction or the observation says an event occurred.
$$L=L_{\text{weighted, masked, multiscale }1}+L_{\text{soft CSI}}$$In words: combine intensity-weighted absolute error over valid pixels and several scales with a differentiable event-score penalty.
The soft event boundary uses a 2 dBZ sigmoid at thresholds of 20 and 35 dBZ. A 63.75 dBZ core counts five times a clear pixel in the intensity weighting. That makes strong echoes matter to the update. The loss is still a stand-in for the score we care about. Source §3.
Training as it ran
The run used batch size 8 on 2 GPUs, a base learning rate of 2e-4 with the head learning five times faster, and 8 epochs. It took 10.87 hours, about 78.5 minutes per epoch. Source §3.
Scored on 24 of the 441 validation samples, chosen across 17 days, the model’s CSI20@30 is 0.695 with interval [0.648, 0.737], against 0.638 for S-PROG. On the full validation split the same checkpoint scores 0.672 against 0.504 for our advection; we have no full-split S-PROG score. The checkpoint rule saved epoch 2, while epoch 4 has the highest logged CSI20. Our advection is the weaker comparison: standard Lucas–Kanade optical-flow extrapolation beats it by about 0.09 CSI, so S-PROG is the comparison that matters. Source §3.
Outputs
The strip compares a saved forecast with its target at the same valid time. It comes from a regional fine-tune, not from the base model scored above.
The flipbook steps through the same validation cases saved after each epoch, with that epoch’s scores. The lead scrubber changes the requested forecast horizon instead. Later training is not later weather. The saved images show the 30-minute lead only; the score table holds the 30- and 60-minute leads.
What it is not
A confidence interval expresses uncertainty in an estimate. The day-block procedure respects shared weather within a day. The winning margin was far smaller than the uncertainty scale, so picking the peak epoch establishes no reliable advantage over its neighbors. A later warm-started run on Hawaii stayed flat, 0.669, 0.667 and 0.668 over three epochs, and changes that small, inside the day-block uncertainty, show no improvement. Source §3.
Before comparing two forecast images, check their starting time and requested lead. A later observation looks like a better forecast when all it did was see more of the event. Changing a reflectivity threshold changes which pixels count as events, so the threshold belongs in the score label and in the comparison protocol.
Where this shows up when you train
Choose a stopping rule that measures the behavior you need, and look at event frequency as well as overlap. Keep the baseline, the per-epoch scores, the region, and the forecast lead visible. Otherwise a convincing curve quietly answers a different question.
Worked example
In a toy map with 6 hits, 2 misses, and 1 false alarm, $\mathrm{CSI}=6/(6+2+1)=6/9$, or two thirds. Adding correctly dry pixels changes neither the numerator nor the denominator. Entirely dry pairs need a scoring convention because that denominator is zero. Source §7.
Run this yourself
trainkit nowcast samples/blobs.npz --output nowcast-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 model scored a CSI of 0.968 on the synthetic blobs at an intensity threshold of 0.5, against 0.587 for the baseline, and trained in about 20 seconds. Source §6.
The gain is $0.968-0.587=0.381$. Against the stronger constant-velocity forecast, which scored 0.843, it is $0.968-0.843=0.125$. We ran it once, so there is no interval on either 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
- Recent radar and motion provide the input; infrared does not enter this run.
- CSI discards correctly dry pixels and must be read with its threshold and lead.
- S-PROG is the comparison that matters; our own advection is a weak baseline.
- A tiny winning checkpoint margin inside the uncertainty band is not reliable progress.
Where we’re headed
The same questions carry into Doppler Everywhere.