Wet-Biased Data Pipelines Mislead the Training and Evaluation of Generative Precipitation Nowcasting

Christoph Metzl; Stefan Kapsreiter; Kianusch Vahid Yousefnia; Virginia Poli; Richard Müller; Frank Krüger; Tobias Bölle

cs.LGOct 19, 2026

Wet-Biased Data Pipelines Mislead the Training and Evaluation of Generative Precipitation Nowcasting

Authors: Christoph Metzl, Stefan Kapsreiter, Kianusch Vahid Yousefnia, Virginia Poli, Richard Müller, Frank Krüger, Tobias Bölle

Organizations: Deutsches Zentrum für Luft- und Raumfahrt, Institut für Physik der Atmosphäre, Oberpfaffenhofen, Germany · Wismar University of Applied Sciences, Wismar, Germany · Arpae Emilia-Romagna, Hydro-Meteo-Climate Service (SIMC), Bologna, Italy · German Weather Service, Offenbach, Germany

Abstract

Generative models for precipitation nowcasting are trained and evaluated predominantly on wet-biased datasets, which over-represent precipitation events to emphasize rare extremes. Selecting test cases based on target observations is known to favor forecasts that match the selected distribution rather than the intended deployment distribution. We show analytically how this target-dependent sampling also shifts the conditional distribution learned by generative models. When training and test sampling match, even an ideal model learns precisely the distribution favored by evaluation, allowing the evaluation bias to conceal the training bias. Using a fixed satellite-to-radar diffusion model, we demonstrate that target-dependent test sampling can reverse the calibration assessment. The model appears under-predictive on a wet-biased test set but over-predictive on a deployment-representative test set. This result is replicated outside the model's training domain after a seven-year gap. Utilizing classifier-free guidance as an example, we demonstrate how the contradictory calibration assessment can misdirect downstream model development.

Figures & tables

Study	Model	Training sampling	Test sampling	Deployment recovery	Probabilistic calibration
DGMR [ 36 ]	GAN	Importance sampling	Stronger sampling	Possible	Reliability (IPW); rank histogram (shifted)
NowcastNet [ 19 ]	GAN	Importance sampling	Stronger sampling and uniform sampling	Possible / uniform set	—
Netherlands DGMR [ 33 ]	GAN	Deterministic filter	Equivalent to training	No	—
DiffCast [ 23 ]	Diffusion	Event-curated / deterministic filter	Equivalent to training	No	—
CasCast [ 22 ]	Diffusion	Event-curated / deterministic filter	Equivalent to training	No	—
LDCast [ 21 ]	Diffusion	Intensity-balanced	Equivalent to training	Possible	Rank histogram (shifted)

Table 1: Dataset sampling and calibration evaluation practices in a sample of influential and recent generative precipitation nowcasting studies. Additional work is surveyed in Supplementary Table 1 . Sampling implications are discussed in Section 3.1 , calibration practices are surveyed in Section 3.2 and their implications discussed in Section 3.3 . Deployment recovery indicates if the estimation of the verification scores on the deployment distribution by inverse probability weighting (IPW) is possible (cf. Section 3.2 ).

Figure 1: Importance sampling creates distinct deployment, training, and test distributions and shifts the empirical training data. a , The climatological deployment distribution

p_{\mathrm{clim}}(x,y)

is shifted into training and test distributions by their respective target-dependent acceptance probabilities,

a_{\mathrm{train}}(y)

and

a_{\mathrm{test}}(y)

. b , Distribution of the acceptance probability

a(y)

over all candidate crops from the 2015–2018 training period. c , Distribution of the same acceptance probability after sampling, evaluated over the crops retained for model training. Histograms in b and c show the percentage of crops per bin on a logarithmic scale. d , Normalized pooled radar-reflectivity distributions for all pixels in the complete training period and for the importance sampled crops actually used to train the model.

Metric	Threshold	Uniform test set	Importance sampled test set	IPW estimate from importance set
Germany
Bias	10 dBZ	1.541 [1.411, 1.705]	0.688 [0.666, 0.715]	1.540 [1.416, 1.697]
	30 dBZ	0.747 [0.643, 0.894]	0.245 [0.209, 0.290]	0.735 [0.635, 0.877]
	38 dBZ	0.687 [0.580, 0.856]	0.233 [0.192, 0.287]	0.667 [0.568, 0.820]
CSI	10 dBZ	0.198 [0.182, 0.213]	0.363 [0.350, 0.376]	0.198 [0.182, 0.213]
	30 dBZ	0.036 [0.028, 0.043]	0.039 [0.032, 0.047]	0.035 [0.028, 0.043]

Table 2: Verification of the same nowcasting model on uniform and importance sampled (wet-biased) test sets over Germany and Italy. The IPW column re-estimates performance under the uniform sampling distribution from the importance sampled set using Horvitz-Thompson totals and Hájek-type ratios. Values in brackets are 95% confidence intervals from a 10,000-replicate block bootstrap with seven-day blocks. Categorical scores and fractions skill score (FSS) are member averaged and aggregated over all lead times. Italian thresholds are the Marshall-Palmer equivalents of the German reflectivity thresholds (0.10, 2.61, and 9.55 mm h -1 , respectively). Pooled CRPS is reported in dBZ for Germany and mm h -1 for Italy and is therefore not comparable across domains.

Figure 2: Rank histograms for the same 20-member forecast ensemble evaluated on uniform and importance sampled (wet-biased) test sets over Germany and Italy. Error bars and the reported mean-rank-bias intervals are

95\text{\,}\mathrm{\%}

confidence intervals (CI) from a 10,000-replicate block bootstrap with seven-day blocks. The signed normalized mean-rank bias ranges from

-1

(all observations below the ensemble) to

+1

(all observations above the ensemble), with zero indicating a centered histogram. The dashed line denotes the relative frequency expected under perfect calibration. Histograms aggregate all lead times.

Figure 3: Reliability diagrams over Germany (top row) and Italy (bottom row) for threshold-matched exceedance levels. Columns show 10, 30, and 38 dBZ over Germany and their Marshall-Palmer equivalents of 0.10, 2.61, and 9.55 mm h -1 over Italy. The same model is evaluated on the uniform test set, on the raw importance sampled (wet-biased) test set, and using an IPW (Horvitz-Thompson/Hájek) estimate from the importance sampled set. Shading denotes 95% confidence intervals from a 10,000-replicate block bootstrap with seven-day blocks. Most intervals are narrower than the markers and are consequently obscured. An interval is omitted where fewer than ten seven-day blocks contribute forecasts to the corresponding probability bin. Marker area follows the logarithm of the corresponding refinement-bin count using one common scale across all panels and estimates. The diagonal denotes perfect reliability, and points aggregate all lead times.

Metric	Threshold	Guidance scale 1	Guidance scale 2	Guidance scale 4
Bias	10 dBZ	1.541 [1.411, 1.705]	1.694 [1.563, 1.855]	2.056 [1.887, 2.268]
	30 dBZ	0.747 [0.643, 0.894]	1.685 [1.451, 2.020]	3.761 [3.233, 4.527]
	38 dBZ	0.687 [0.580, 0.856]	2.137 [1.804, 2.695]	6.467 [5.354, 8.409]
CSI	10 dBZ	0.198 [0.182, 0.213]	0.241 [0.222, 0.259]	0.249 [0.227, 0.269]
	30 dBZ	0.036 [0.028, 0.043]	0.064 [0.051, 0.076]	0.067 [0.054, 0.080]
	38 dBZ	0.017 [0.013, 0.022]	0.035 [0.026, 0.043]	0.031 [0.023, 0.039]

Table 3: Verification of classifier-free guidance scales 1, 2, and 4 on the German uniform test set. Values in brackets are 95% confidence intervals from a 10,000-replicate block bootstrap with seven-day blocks. Categorical scores and fractions skill score (FSS) are member averaged and aggregated over all lead times; CRPS is reported in dBZ. Guidance scale 1 is repeated from Table 2 to make the guidance comparison self-contained.

Figure 4: Calibration of classifier-free guidance scales 1, 2, and 4 on the German uniform test set. Each row shows one guidance scale. The left column contains rank histograms aggregated over all lead times. Error bars and confidence intervals are estimated the same way as in Fig. 2 . The right column contains reliability diagrams for exceedance of 10, 30, and 38 dBZ. Confidence intervals and refinement-bin counts are depicted analogously to Fig. 3 . The dashed lines denote perfect calibration.

Figure 5: German light-rain case on 24 July 2020, 04:00 UTC and a lead time of 30 minutes. Each panel shows the same $128\times 128$ -pixel crop at $1\text{\,}\mathrm{km}$ spatial resolution. a,b , The same paired ensemble member at guidance scales $s=1$ and $s=2$ ; c , ground truth. d,e , Exceedance probabilities $p_{s=1}$ and $p_{s=2}$ , defined at each pixel as the fraction of the 20 ensemble members reaching or exceeding 10 dBZ at the displayed lead time. f , The pixelwise probability difference $\Delta p=p_{s=2}-p_{s=1}$ . Aggregated over all members, lead times, and pixels in this case, the observed exceedance fraction is 3.67%, compared with forecast fractions of 23.58% at $s=1$ and 26.87% at $s=2$ . The corresponding CSI changes only from 0.044 to 0.047, so the additional 3.29 percentage points of forecast event area yield little categorical benefit while strengthening the already excessive light-rain coverage.

Figure 6: Italian convective case on 13 July 2025, 14:00 UTC and a lead time of 30 minutes. Each panel shows the same $128\times 128$ -pixel crop at $1\text{\,}\mathrm{km}$ spatial resolution. a,b , The same paired ensemble member at guidance scales $s=1$ and $s=2$ ; c , ground truth. d,e , Exceedance probabilities $p_{s=1}$ and $p_{s=2}$ , defined at each pixel as the fraction of the 20 ensemble members reaching or exceeding 9.55 mm h -1 at the displayed lead time. f , The pixelwise probability difference $\Delta p=p_{s=2}-p_{s=1}$ . Aggregated over all members, lead times, and pixels in this case, increasing the guidance scale from $s=1$ to $s=2$ raises CSI from 0.057 to 0.168 and moves the forecast exceedance fraction from 0.55% to 2.03%, compared with 1.59% in the observations. At the same time, the median member maximum increases from 94.67 to 322.04 mm h -1 , illustrating that the improved localization and categorical overlap are accompanied by excessive peak intensities and additional forecast event area.

Figure 7: Geolocated input domains and crop construction. a,b , German RADOLAN WX radar reflectivity in dBZ and the corresponding Meteosat-10 SEVIRI high-resolution visible (HRV) channel at 1 June 2016, 11:45 UTC, near the peak of the Simbach flooding; c,d , Italian SRI radar rain rate in mm h -1 and the corresponding HRV channel at 16 May 2023, 06:00 UTC, during the Emilia–Romagna flooding. Black outlines in the upper-left corner of each radar panel show one solid $128\times 128$ -pixel reference crop and its dashed right and lower neighbours.

Study	Model	Training sampling	Test sampling	Deployment recovery	Probabilistic calibration
LangPrecip [ 1 ]	Flow matching	Event-curated; filtering unclear	Equivalent to training	Unclear	—
PostCast [ 2 ]	Diffusion	Mixed; dataset-inherited	Random 500 per dataset from test splits	Mixed	—
PreDiff [ 3 ]	Diffusion	Event-curated	Equivalent to training	No	—
DYffCast [ 4 ]	Diffusion	Unclear	Unclear	Unclear	Spread–skill only
DuoCast [ 5 ]	Diffusion	Event-curated / deterministic filtering	Equivalent to training	No	—
MoCast+ [ 6 ]	Diffusion	Event-curated / deterministic filtering	Equivalent to training	No	—

Supplementary Table 1: Dataset sampling and calibration evaluation practices in additional generative precipitation nowcasting studies. Studies included in Table 1 are omitted.

Parameter	Symbol	Value
Saturation scale	$s$	1 mm h -1
Rain contribution coefficient	$m$	2
Minimum acceptance probability	$q_{\min}$	0.004
Time steps	$T_{y}$	3
Crop height	$H$	128 pixels
Crop width	$W$	128 pixels

Supplementary Table 2 : Parameters of the target-dependent importance sampling scheme used to construct the training set. The normalization size

D

is the number of radar values in one target crop.

Supplementary Fig. 1 : Sensitivity of the pooled training-target distribution to the effective acceptance coefficients. Curves show normalized radar-reflectivity frequencies across all target pixels retained under the training configuration in Supplementary Table 2 and two sensitivity settings. The lower-acceptance setting scales

m

and

q_{\min}

0.05

relative to the training configuration, whereas the higher-acceptance setting scales both by

5

; all other parameters are unchanged. Higher acceptance causes a larger fraction of candidate crops to reach unit acceptance probability, reduces the relative enrichment of rainy crops, and shifts the retained distribution toward no-rain values.

Setting	No rain ( $<10$ dBZ)	Light ( $10$ – $30$ dBZ)	Moderate ( $30$ – $38$ dBZ)	Strong ( $38$ – $48$ dBZ)	Extreme ( $\geq 48$ dBZ)
Lower-acceptance sensitivity ( $0.05\times$ )	56.251	35.519	6.852	1.281	0.098
Training configuration	57.788	34.767	6.188	1.160	0.097
Higher-acceptance sensitivity ( $5\times$ )	72.465	23.538	3.270	0.655	0.071

Supplementary Table 3 : Radar-reflectivity class fractions in the retained training crops under the training configuration and two sensitivity settings. Values are percentages of all pooled target pixels and sum to 100% within each row. The lower- and higher-acceptance settings multiply both

m

and

q_{\min}

0.05

and

5

, respectively, relative to Supplementary Table 2 , while holding

s

T_{y}

H

, and

W

fixed. Class boundaries are left-inclusive and right-exclusive except for the open outer bounds.

Quantity	Germany (2015–2020)	Italy (2025)
Timestamps on the 15-min grid	105,984	17,664
Globally invalid timestamps	807	30
missing satellite source files	690	16
missing satellite values	35	14
missing radar files	82	0
Crop-timestamps	9,644,544	5,299,200

Supplementary Table 4 : Effect of quality control on sample count for both domains. Timestamps refer to the 15-min grid of the May–October seasons. A timestamp is globally invalid if the radar file is missing, if at least one of the SEVIRI channel files is missing so that the scene cannot be gridded, or if a gridded channel contains missing values. Crop-timestamps are timestamps multiplied by the number of lattice crops (91 over Germany, 300 over Italy). A crop-timestamp is masked if its timestamp is globally invalid or if more than 360 of its radar pixels are missing. Candidate sequences are six-frame samples for which all frames are valid and fall within May–October.

Split (season)	Candidates	Importance sampled	Uniform	Acceptance rate
Training (2015–2018)	6,221,943	425,713	—	6.84%
Validation (2019)	1,568,930	125,765	125,765	8.02%
Test Germany (2020)	1,543,127	122,644	122,644	7.95%
Test Italy (2025)	1,967,513	95,111	95,111	4.83%

Supplementary Table 5 : Sizes of all datasets used in this study. Importance sampled sets use Eq. ( 6 ), while uniform sets are simple random samples without replacement of the same size as the corresponding importance sampled set. The acceptance rate is the size of the importance sampled set relative to the candidate pool. Only the uniform validation set is used for checkpoint selection.

Supplementary Fig. 2 : Schematic of one EDM training step in the notation of Section 5.2 .

Supplementary Fig. 3 : High-level architecture of the conditional 2D U-Net

F_{\theta}

. The three noisy radar frames are concatenated with the full-resolution satellite features and the temporal dimension is folded into the channel dimension. Encoder and decoder operate on four resolution levels from

128\times 128

16\times 16

pixels with 128 to 1024 channels and two residual blocks per level; the bottleneck additionally contains an 8-head spatial self-attention block. Downsampling uses strided

3\times 3

convolutions and upsampling nearest-neighbour resizing followed by a

3\times 3

convolution. Skip connections are concatenated and reduced by a

1\times 1

projection. Multi-scale satellite conditioning (green arrows) is added after each downsampling and before each upsampling. A final

1\times 1

convolution yields the three output frames.

Panel of Figure 19 — Supplementary Fig. 4 : Conditioning components of the conditional diffusion model. (a) Satellite conditioning encoder. Each of the three input frames with four channels passes through two convolution blocks (convolution, group normalization, SiLU) with 32 and 16 channels. A 16-dimensional sine–cosine encoding of the time of day of each frame, followed by a linear layer, is added to the per-frame features. After restoring the temporal axis, a $3\times 3\times 3$ convolution block fuses the frames into 64 feature channels per time step. This representation is concatenated with the noisy radar at the U-Net input (full-resolution conditioning) and, after spatial pooling, folding time into channels, and a $1\times 1$ projection, added at the $64^{2}$ , $32^{2}$ , and $16^{2}$ levels (multi-scale conditioning). (b) Conditioned residual block. The noise level $c_{\mathrm{noise}}$ is mapped by a Fourier embedding, a linear layer, and SiLU to channel-wise scale and shift parameters that modulate the group-normalized activations (adaptive GroupNorm, AdaGN). The block consists of two $3\times 3$ convolutions with group normalization and SiLU and a residual connection that is the identity or a $1\times 1$ convolution when the channel count changes.

Hyperparameter	Value
Data and preprocessing
Satellite predictors	HRV, IR_108, VIS008, WV_062
Input/target sequence	3 input frames ( $-30,-15,0$ min); 3 target frames ( $0,15,30$ min)
Temporal resolution	15 min
Crop size and stride	$128\times 128$ pixels; 64-pixel stride (50% overlap)
Satellite standardization $(\mu,\sigma)$	HRV $(11.3462,14.2666)$ ; IR_108 $(270.7699,16.8125)$ ; VIS008 $(12.9961,15.3883)$ ; WV_062 $(232.0065,4.3216)$

Supplementary Table 6 : Hyperparameters used for model training and ensemble inference. The baseline and IPW models differ only in the per-sample loss multiplier. The model was implemented in PyTorch 2.5.1 [ 31 ] with PyTorch Lightning. Values labelled as PyTorch defaults are stated explicitly. The epoch is zero indexed. Training and inference are seeded, but bitwise reproducibility is not guaranteed because deterministic kernels were not enforced and non-deterministic GPU algorithms were not disabled. All computation is in 32-bit floating point.

Supplementary Fig. 5 : Radially averaged power spectral density (RAPSD) of the model forecasts and corresponding ground-truth fields over Germany (top row) and Italy (bottom row), evaluated on the uniform (left column) and importance sampled (right column) test sets. Curves aggregate all forecast lead times. Forecast spectra use the color of the respective sampling protocol, while ground truth is shown by the charcoal dashed curves. Both axes are logarithmic, and spatial frequency is given in pixel -1 . The absolute spectral densities should not be compared across domains because the German fields contain radar reflectivity in dBZ, whereas the Italian fields contain precipitation rates in mm h -1 . Notably, the model reproduces the RAPSD almost perfectly when evaluated on the importance sampled German set. However, the model lies consistently above the ground truth on the uniform set. Furthermore, the RAPSD does not generalize well to the Italian domain on either set.

Supplementary Fig. 6 : Critical success index (CSI; top row) and fractions skill score at a neighborhood scale of 128 pixels (

\mathrm{FSS}_{128}

; bottom row) as functions of lead time on the uniform German and Italian test sets. Columns correspond to threshold-matched intensity levels of 10, 30, and 38 dBZ over Germany and 0.10, 2.61, and 9.55 mm h -1 over Italy, respectively. Scores are member averaged; solid blue curves with circles denote Germany and dashed magenta curves with squares denote Italy. Across the available 0–30-minute range, both scores are nearly invariant with lead time, apart from a modest decrease in CSI at the highest threshold. This weak temporal degradation is consistent with the satellite-to-radar conversion being the dominant difficulty over these short lead times, rather than forecast extrapolation alone.

Metric	Threshold	Uniform test set	Importance sampled test set	IPW estimate from importance set
Germany
POD	10 dBZ	0.420 [0.411, 0.431]	0.450 [0.439, 0.460]	0.421 [0.411, 0.431]
	30 dBZ	0.060 [0.050, 0.072]	0.047 [0.038, 0.058]	0.059 [0.048, 0.072]
	38 dBZ	0.029 [0.022, 0.036]	0.019 [0.013, 0.024]	0.028 [0.021, 0.035]
FAR	10 dBZ	0.727 [0.700, 0.756]	0.347 [0.322, 0.375]	0.727 [0.701, 0.755]
	30 dBZ	0.919 [0.900, 0.939]	0.809 [0.785, 0.840]	0.919 [0.901, 0.938]

Supplementary Table 7 : Complementary verification scores for Table 2 . Definitions, aggregation, threshold matching, evaluation protocols, and confidence intervals are as in Table 2 . FSS is shown at the two neighborhood scales omitted from the main table. CRPS is reported in dBZ for Germany and mm h -1 for Italy and is not comparable across domains.

Metric	Threshold	Guidance scale 1	Guidance scale 2	Guidance scale 4
POD	10 dBZ	0.420 [0.411, 0.431]	0.523 [0.505, 0.541]	0.609 [0.589, 0.631]
	30 dBZ	0.060 [0.050, 0.072]	0.161 [0.135, 0.190]	0.301 [0.264, 0.338]
	38 dBZ	0.029 [0.022, 0.036]	0.107 [0.087, 0.125]	0.226 [0.196, 0.252]
FAR	10 dBZ	0.727 [0.700, 0.756]	0.692 [0.664, 0.720]	0.704 [0.677, 0.733]
	30 dBZ	0.919 [0.900, 0.939]	0.904 [0.885, 0.925]	0.920 [0.904, 0.937]
	38 dBZ	0.958 [0.946, 0.972]	0.950 [0.937, 0.964]	0.965 [0.955, 0.975]

Supplementary Table 8 : Complementary verification scores for the classifier-free guidance comparison in Table 3 . The evaluation set, aggregation, and confidence intervals are as in Table 3 . FSS is shown at the two neighborhood scales omitted from the main table; CRPS is reported in dBZ.

Metric	Threshold	Baseline	IPW-trained
Bias	10 dBZ	1.541 [1.411, 1.705]	1.416 [1.295, 1.570]
	30 dBZ	0.747 [0.643, 0.894]	0.670 [0.578, 0.804]
	38 dBZ	0.687 [0.580, 0.856]	0.673 [0.565, 0.856]
POD	10 dBZ	0.420 [0.411, 0.431]	0.329 [0.323, 0.337]
	30 dBZ	0.060 [0.050, 0.072]	0.031 [0.026, 0.036]
	38 dBZ	0.029 [0.022, 0.036]	0.012 [0.009, 0.015]

Supplementary Table 9 : Verification of the baseline and inverse-probability-weighted (IPW) trained models on the German uniform test set at guidance scale 1. Values in brackets are 95% confidence intervals from a 10,000-replicate block bootstrap with seven-day blocks. Categorical scores and fractions skill score (FSS) are member averaged and aggregated over all lead times; Pooled CRPS and CRPS are reported in dBZ. The baseline is repeated from Table 2 to make the model comparison self-contained.

Supplementary Fig. 7 : Calibration of the baseline and inverse-probability-weighted (IPW) trained models on the German uniform test set at guidance scale 1. Both models are trained on the same importance sampled training set; the IPW-trained model weights its loss by the inverse sampling acceptance probability. Each row shows one model. The left column contains rank histograms aggregated over all lead times; error bars and the reported signed normalized mean-rank-bias intervals are 95% confidence intervals from a 10,000-replicate block bootstrap with seven-day blocks. The dashed horizontal line denotes the relative frequency expected under perfect calibration. The right column contains reliability diagrams for exceedance of 10, 30, and 38 dBZ. Shading denotes the corresponding 95% block-bootstrap confidence intervals; intervals are omitted where fewer than ten seven-day blocks contribute to a probability bin. Marker area follows the logarithm of the refinement-bin count using one common scale across panels and models, and the diagonal denotes perfect reliability. IPW training moves the rank histogram and mean-rank bias toward perfect calibration, although residual overprediction remains.

Supplementary Fig. 8 : Latest satellite inputs ( $t=0$ ) for the German convective case in Supplementary Fig. 9 : a , high-resolution visible (HRV); b , 10.8- $\mu$ m infrared; c , 0.8- $\mu$ m visible; and d , 6.2- $\mu$ m water vapour. Visible channels show reflectance and thermal channels brightness temperature.

Supplementary Fig. 9 : German convective case on 10 May 2020, 15:15 UTC and a lead time of 30 minutes. Each panel shows the same $128\times 128$ -pixel crop at $1\text{\,}\mathrm{km}$ spatial resolution. a,b , The same paired ensemble member at guidance scales $s=1$ and $s=2$ ; c , ground truth. d,e , Exceedance probabilities $p_{s=1}$ and $p_{s=2}$ , defined at each pixel as the fraction of the 20 ensemble members reaching or exceeding 38 dBZ at the displayed lead time. f , The pixelwise probability difference $\Delta p=p_{s=2}-p_{s=1}$ . Aggregated over all members, lead times, and pixels in this case, increasing the guidance scale from $s=1$ to $s=2$ raises CSI from 0.060 to 0.132 and the forecast exceedance fraction from 0.66% to 2.28%, compared with 1.36% in the observations. The median member maximum increases from 54.08 to 59.71 dBZ.

Supplementary Fig. 10 : Latest satellite inputs ( $t=0$ ) for the Italian light-rain case in Supplementary Fig. 11 . Panels and units are defined as in Supplementary Fig. 8 ; the near-zero visible-channel reflectances are expected at 01:15 UTC.

Supplementary Fig. 11 : Italian light-rain case on 20 June 2025, 01:15 UTC and a lead time of 30 minutes. Crop dimensions, spatial resolution, panel definitions, and guidance scales are as in Supplementary Fig. 9 , using an exceedance threshold of 0.10 mm h -1 . Aggregated over all members, lead times, and pixels in this case, the observed exceedance fraction is 2.86%, whereas the forecast fraction increases from 12.87% at $s=1$ to 17.88% at $s=2$ . CSI rises only from 0.054 to 0.064, so the additional 5.01 percentage points of forecast event area primarily amplify false light-rain coverage.

Precipitation Nowcasting

Precipitation nowcasting predicts rainfall over minutes to hours from radar, satellite, station, and numerical weather data.

See all 40 papers

Oct 9, 2026

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Accurate daily precipitation forecasts are important for weather-sensitive decisions. Although recent Weather Foundation Models (WFMs) have achieved significant progress, they are trained on reanalysis data, which exhibit regional and intensity-dependent biases. Since retraining WFMs on observation-based targets is costly, we study Dynamics-Precipitation Decoupling, a plug-and-play framework that uses a separately trained generator to map non-precipitation forecasts from a frozen weather model to observation-aligned daily precipitation. Its key challenges are to capture heterogeneous rainfall distributions and align daily accumulations with atmospheric signals that evolve across space and time. To address them, we train a diffusion generator conditional on atmospheric states and satellite-gauge estimates of precipitation. A multi-scale recurrent encoder summarizes atmospheric evolution, while cross-attention aligns atmospheric and precipitation features during denoising. Extensive evaluation of various backbones, including Pangu-Weather, GraphCast, and HRES, on global and regional forecasting shows improved daily precipitation accuracy and event detection, with better representation of localized accumulation patterns. Global ensemble evaluation shows lower Brier scores than ENS and training-from-scratch GenCast at all reported lead times, with lower CRPS in most comparisons. Remarkably, a heavy-rainfall case study using 2,288 stations in China further supports improved daily accumulation patterns and event detection against ground observations.

Guowen Li, Yang Liu, Mengxuan Chen +11

Tsinghua University · The Chinese University of Hong Kong · The Hong Kong University of S & T +5

Precipitation Forecasting Weather Forecasting Probabilistic Forecasting Conditional Diffusion Models

Ask PDF

Oct 8, 2026

Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion

Generative diffusion models are well-suited for probabilistic precipitation nowcasting, but existing approaches often rely on separately trained compression or deterministic forecasting components and remain costly at inference due to iterative denoising. We introduce Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model which addresses both issues by forecasting directly in radar space and enabling few-step generation. Radar-space modeling greatly simplifies training and inference and eliminates uncertainty arising from lossy compression. JWS combines Masked Asynchronous Diffusion, a timestep-sampling scheme that preserves clean context while adapting diffusion training to high-dimensional spatio-temporal data, with a simple scoring-rule objective that aligns training with probabilistic forecasting and unlocks few-step generation. On the SEVIR and MeteoNet benchmarks, JWS achieves state-of-the-art probabilistic forecasting performance at reduced training and inference cost. Even our smallest model remains competitive using substantially fewer parameters and more than 17x faster inference.

Jannik Wiese, Johannes Schusterbauer, Tommaso Martorella +1

Precipitation Nowcasting Probabilistic Forecasting Spatiotemporal Forecasting Diffusion Models Diffusion Model Inference Acceleration Masked Diffusion Models

Ask PDF

Sep 27, 2026

Physics-Guided Flow-Map Matching for Precipitation Nowcasting

Precipitation nowcasting, generating future radar fields from past observations, is critical for flood warning and disaster response. It is also a demanding benchmark for spatiotemporal generative modeling, with chaotic dynamics, heavy-tailed intensities, and rare high-intensity structures that matter most. Deterministic models minimize a pixel loss and are driven toward the conditional mean, which blurs exactly those structures, while generative models that add a stochastic residual on top of a deterministic backbone inherit the same blur. We propose Physics-Guided Flow-Map Matching (PG-FMM), a conditional flow-map model that decouples predictable advection from uncertain small-scale detail. A frozen Lagrangian advection prior transports the radar field and supplies an explicit motion forecast, and a flow-map generative head, conditioned on the past frames and the prior rollout rather than summed onto it, produces sharp stochastic detail in four sampling steps. The prior serves only as guidance, so the head replaces blurred structure instead of inheriting it. Extensive experiments on four radar benchmarks show that PG-FMM outperforms state-of-the-art methods on 18 of 24 metrics, with the largest gains at heavy-rain thresholds, where the critical success index improves by up to 58.9%. The project page can be found at https://neurogica.github.io/PG-FMM.

Shunya Nagashima, Takumi Bannai, Makoto Misaizu +3

Neurogica Inc., Tokyo, Japan · LTS, Inc., Tokyo, Japan · ME-Lab Japan, Inc., Tokyo, Japan +1

Precipitation Nowcasting Conditional Generative Modeling Physics-Informed Generative Modeling Flow Map Learning

Ask PDF

Abstract

Precipitation Nowcasting

Explore similar work

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion

Physics-Guided Flow-Map Matching for Precipitation Nowcasting