Short-range rain forecasts made from radar, called nowcasts, are now mostly produced by neural networks that learn to extrapolate the last hour of radar images. They do well at moving existing storms around. They do poorly at the event forecasters care most about: a storm that grows or dies faster than its recent history suggests. The atmospheric science literature explains why. Rain intensity is set by two things. The first is where the storm is carried, which is a question of motion. The second is how much moisture and instability the storm meets along its path, which is a question of the local water and energy budget. Radar images contain a great deal of information about the first and very little about the second.
We propose a nowcaster that writes the forecast as "rain is carried along a learned motion field, plus a growth-or-decay term." The motion is learned only from fast radar frames. The growth-or-decay term is the only part of the model allowed to see the slow environmental state: moisture flux convergence, precipitable water, and near-surface temperature tendencies. The claim is that this split matters for skill on growth and decay events, not just for interpretability. Because the growth term is explicit, the trained model also becomes a measuring instrument. It can test, on real observations, two open questions: whether organized convection lengthens storm lifetimes or raises peak rates, and why rainfall's sensitivity to temperature differs so much from region to region.
Core Claim
Write radar-derived rain intensity as $I(\mathbf{x},t)$. The model evolves it as
$$\frac{\partial I}{\partial t} + \mathbf{v}\cdot\nabla I = S,$$
where $\mathbf{v}(\mathbf{x},t)$ is a motion field predicted from recent radar frames, and $S(\mathbf{x},t)$ is a source term that adds or removes rain. $S$ is predicted from the current intensity and from a slow environmental stream $\mathbf{e}$ that contains thermodynamic and moisture-transport fields. The architectural rule that makes this more than a restatement of advection is information routing. $\mathbf{e}$ reaches $S$ and nothing else. $\mathbf{v}$ sees radar only, or radar plus steering winds in an ablation.
Skilful nowcasting of extreme precipitation with NowcastNet supplies the method. Its evolution network already solves a differentiable continuity equation with a motion field and an intensity residual, and a generative network then adds fine detail. In NowcastNet, however, both the motion field and the residual are inferred from radar alone. A storm forming in clear air, or decaying while it looks healthy, is therefore invisible to it. Our change is to give the residual a physically meaningful input path while keeping motion estimation clean.
Two papers motivate the claim. Implicit learning of convective organization explains precipitation stochasticity shows that coarse thermodynamic variables alone explain extreme rain poorly ($R^2\approx0.45$). Adding a learned organization summary raises this to $R^2\approx0.9$, and that summary is highly predictable from its own past. This tells us the source term needs both environment and storm-state inputs. The role of cyclonic activity in tropical temperature-rainfall scaling shows that local rainfall–temperature scaling turns negative over about a third of tropical ocean locations, driven by pre-rainfall cooling ($r=-0.987$). This tells us the source term's response to temperature cannot be hardwired as a fixed thermodynamic increase of about 7% per kelvin, and must be learned conditionally.
Research Program
The fast stream follows the asymmetric design of SlowFast Networks for Video Recognition. SlowFast splits a video model into a high-frame-rate, low-capacity "motion" pathway and a low-frame-rate, high-capacity "content" pathway, with the fast pathway costing about 20% of the compute. We reuse the asymmetry but change what defines the slow pathway: physical variables on hourly reanalysis cadence rather than subsampled frames of the same video. The fast pathway ingests 5-minute radar and outputs $\mathbf{v}$. For motion we borrow two ideas from learned object tracking:
- a bidirectional state-space sequence encoder over past motion, as in MambaTrack;
- a direction-consistency loss on predicted displacement, as in ETTrack's momentum correction (ETTrack).
Both papers show learned predictors beating constant-velocity Kalman filters on curved, erratic trajectories. They track only geometry, though, and our point is to pair that kind of motion model with an explicit, separately conditioned intensity evolution.
The slow stream is built from ERA5 fields at or before forecast issue time. Future reanalysis is never used, to avoid leakage. The fields are:
- vertically integrated moisture flux divergence;
- precipitable water;
- near-surface temperature;
- its one-hour tendency, a proxy for the pre-rain cooling Traxl et al. identify.
The context window for these fields is larger than the radar patch. Analysis of the Atmospheric Water Budget for Elucidating the Spatial Scale of Precipitation Changes Under Climate Change shows that below roughly 3,500 km, precipitation is governed by moisture transport rather than local evaporation. A 384 km radar patch is therefore budget-open by construction. We do not expect 3,500 km of context to be needed at a three-hour horizon, because that scale describes climate-mean closure, not forecast-horizon transport. Our hypothesis is that the useful context radius grows with lead time roughly as typical steering speed times lead time, since that is the distance moisture can travel into the patch before verification. Measuring this curve is a direct, falsifiable link between the scaling literature and nowcast design.
Organization enters $S$ through a small encoder over the current radar field that summarizes clustering, in the spirit of the learned "org" variable from the organization paper, with the nearest-neighbor clustering index $I_{\text{org}}$ computed alongside for interpretation. A generative refinement stage on top of the evolution output handles the fine-scale texture that deterministic losses blur.
The scientific analyses use the trained model as an instrument.
Storm-object budgets. We link cells across frames with the learned motion field and integrate $S$ along each cell's path. This splits each cell's accumulated rain into lifetime and peak-rate components, so we can ask, in observations, whether higher organization predicts longer lifetime or higher peaks. This is the tension between Significant Amplification of Instantaneous Extreme Precipitation With Convective Self‐Aggregation, which finds microphysically driven increases in instantaneous extremes, and Intensification of daily tropical precipitation extremes from more organized convection, which attributes daily super-CC scaling mostly to duration and notes the organization signal is weaker in satellite observations than in the model. Our analysis is over the continental US rather than tropical oceans, so it tests whether the duration story transfers. It does not settle the tropical case.
Sensitivity maps. We compute partial derivatives of $S$ with respect to temperature, with moisture convergence held fixed and allowed to vary, and map them by region. This asks whether regional heterogeneity in the temperature response disappears once the dynamic term is conditioned on. That is the mechanism Traxl et al. propose for tropical oceans.
Baselines
The naive reference is Lagrangian persistence: today's field advected by optical flow, with no growth or decay. We run it with the pySTEPS-style flow used in the NowcastNet comparisons. Any model that cannot beat it on growth and decay events has learned nothing about intensification. On the Mississippi River Basin setting, FusionCast reports optical flow at 1.598 mm/h RMSE.
The strongest radar-only physical competitor is NowcastNet itself, retrained on our splits with identical inputs except the slow stream. It is the closest architecture, and its residual term is exactly the one we reroute. No reported result was found on Lacuna for NowcastNet on SEVIR, so we report our retrained numbers.
The strongest thermodynamically informed competitor is FusionCast. It fuses GNSS precipitable water with radar through spatial and channel attention gates and uses a NowcastNet forecast as a prior. The reported numbers are RMSE 1.004 mm/h versus NowcastNet's 1.500 mm/h, and heavy-rain CSI (critical success index, the overlap score for threshold exceedance) of 0.213 versus 0.208 at 120 minutes at 4 mm/h. We reproduce its Mississippi River Basin February–June 2023 period. Because the GNSS network may not be available to us, we run FusionCast's gate with ERA5 precipitable water on the same grid our model sees. That makes the comparison about how thermodynamics enters the model, not which sensor supplies it.
The strongest unconstrained video model is PreDiff (PreDiff: Precipitation Nowcasting with Latent Diffusion Models). It reports Fréchet Video Distance of 33.05 on SEVIR 60-minute VIL versus 133.0 for a standard latent diffusion model. We also include exPreCast (Extreme Weather Nowcasting via Local Precipitation Pattern Prediction), a frequency-split deterministic transformer that is efficient on extremes. No numeric SEVIR result for exPreCast was found on Lacuna.
All models get the same training events, the same slow-stream channels where they accept them (concatenated as extra inputs), and an equal hyperparameter search budget in GPU-hours.
The decisive ablation family isolates the routing rule. Variant (a) gives $S$ radar only, which is NowcastNet-like. Variant (b) feeds the environment to both $\mathbf{v}$ and $S$ with no routing constraint. Variant (c) uses the environment only as a multiplicative gate, which is FusionCast-like. Variant (d) shuffles the environment in time across events to destroy its information while preserving its statistics.
Beating the baselines means two things. First, a statistically significant CSI gain over variants (a), (b), and (c) and over PreDiff on growth and decay subsets, by bootstrap over events. Second, no loss relative to NowcastNet on steady-advection cases.
Why not just…?
Why not just concatenate the reanalysis fields as extra input channels to Earthformer or NowcastNet? This is the most serious threat, and variant (b) together with the concatenated baselines tests it directly. FusionCast reports that naive stacking of mismatched radar and moisture inputs produced interference and smoothed fields, but that is one study with one sensor mix. If concatenation matches routed $S$ on growth and decay CSI, we will report that the physical split buys interpretability and not skill, and drop the skill claim.
Why not just use FusionCast's moisture gate? A multiplicative gate can suppress echoes where air is dry, but it cannot create rain where no echo exists, and it does not see moisture convergence or cooling tendencies. The prediction is specific: variant (c) should match our model on decay events and fall behind on initiation and growth events. If it does not, the gate is sufficient.
Why not just use a large diffusion video model and let it learn intensification? PreDiff's sample quality is strong. A pure-radar generator, however, has no access to the environment, so it cannot anticipate growth that leaves no radar precursor. We compare against PreDiff both radar-only and with environment channels concatenated. If the conditioned diffusion model wins, our contribution reduces to the diagnostic analyses, and we will say so.
Why not just track cells with a Kalman filter and regress intensity change on the environment? This object-based alternative is cheap and interpretable. It inherits the constant-velocity failures shown on DanceTrack-style motion, where ByteTrack scores 47.3 HOTA against MambaTrack's 56.8. It also handles merging and splitting poorly. We run it as an extra baseline on the per-cell lifetime and peak metrics, where it is most competitive.
Proposed Experiments
The primary benchmark is SEVIR: 4-hour, 384 km events at 5-minute cadence, with NEXRAD vertically integrated liquid (VIL) and GOES water-vapor imagery. Each event is matched to ERA5 by time and location. The secondary benchmark is the MRMS Mississippi River Basin setting, for the FusionCast comparison. Events are split by year to avoid temporal leakage.
The metrics are:
- CSI and Heidke skill score at the standard SEVIR VIL thresholds, at 30-minute intervals to 3 hours;
- radially averaged power spectra, to detect blurring;
- per-cell errors in lifetime and peak intensity from the tracking analysis.
The headline evaluation reports these separately on three event classes, defined by the sign and size of the change from Lagrangian persistence at verification: growth, decay, and steady advection. Pooled scores hide exactly the cases this proposal targets.
Three checks decide whether the idea holds. First, the routed model beats variants (a), (b), and (c) on growth events but not on steady events. If the gain is uniform across classes, it comes from extra inputs rather than routing. Second, the useful slow-stream context radius, measured by retraining with radii from the patch size up to about 2,000 km, increases with lead time as transport predicts rather than saturating immediately. Third, the time-shuffled environment control (variant d) erases the growth-event gain. On the scientific side, the storm-object budget is reported only if the model's cell lifetimes and peaks match radar-observed ones within the persistence baseline's error. Otherwise the decomposition reflects the model, not the atmosphere.
References
- Skilful nowcasting of extreme precipitation with NowcastNet
- Implicit learning of convective organization explains precipitation stochasticity
- The role of cyclonic activity in tropical temperature-rainfall scaling
- SlowFast Networks for Video Recognition
- MambaTrack: A Simple Baseline for Multiple Object Tracking with State Space Model
- ETTrack: enhanced temporal motion predictor for multi-object tracking
- Analysis of the Atmospheric Water Budget for Elucidating the Spatial Scale of Precipitation Changes Under Climate Change
- Significant Amplification of Instantaneous Extreme Precipitation With Convective Self‐Aggregation
- Intensification of daily tropical precipitation extremes from more organized convection
- FusionCast: Enhancing Precipitation Nowcasting with Asymmetric Cross-Modal Fusion and Future Radar Priors
- PreDiff: Precipitation Nowcasting with Latent Diffusion Models
- Extreme Weather Nowcasting via Local Precipitation Pattern Prediction