DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting

1. Introduction

• Precipitation nowcasting predicts future radar echo sequences from current observations for very short-range rainfall forecasting, including horizons of 0–6 hours.The task is relevant to meteorological science and smart city applications.Source: §1 · p. 1
• Precipitation systems combine deterministic global movement with stochastic local particle growth and decay, and these dynamics are inherently coupled.The passage illustrates this with evolving convection and continuous rainfall when cold and warm air masses cross.Source: §1 · p. 1
• Deterministic methods capture movement but blur and underestimate high-value echoes, whereas probabilistic methods produce realistic detail but often misplace rainfall.The authors attribute these shortcomings to neglected local stochasticity in deterministic models and excessive generative freedom in probabilistic models.Source: §1 · pp. 1–2, Figure 1 · p. 1
• DiffCast decomposes precipitation evolution into a global motion trend and local stochastic residual, modeled by a deterministic backbone and temporal residual diffusion component.The components are trained end-to-end; the diffusion component uses multi-scale temporal features as conditions.Source: §1 · p. 2
• Experiments on four publicly available radar datasets report state-of-the-art precipitation nowcasting results for DiffCast.The authors report that the framework improves the performance of several deterministic backbones.Source: §1 · p. 2

2.1. Spatio-temporal Predictive Models

• Deterministic predictors include recurrent models that generate frames successively and recurrent-free models that decode all future frames at once.Examples include ConvLSTM and ConvGRU for recurrent prediction, and SimVP and Earthformer for recurrent-free prediction.Source: §2.1 · p. 2
• For precipitation nowcasting, deterministic methods can blur and lose high-value echoes because they neglect local stochasticity.Probabilistic predictors model conditional uncertainty using adversarial training or variational autoencoders, but whole-system stochastic modeling can harm accuracy and training stability.Source: §2.1 · p. 2

2.2. Spatio-temporal Diffusion Models

• Spatio-temporal diffusion methods use denoising diffusion models; examples include MCVD, LDCast, and PreDiff, which add different forms of conditioning.MCVD uses random frame masks, LDCast uses channel-based conditional latent diffusion, and PreDiff adds a knowledge-control network.Source: §2.2 · pp. 2–3
• Diffusion predictors can generate realistic details, but modeling the whole precipitation system stochastically introduces freedom that can mismatch predicted rainfall positions.The cited methods include MCVD, LDCast, and PreDiff.Source: §2.2 · p. 3

3.1. Task Definition

• Given L_in initial frames x, the task models p(y|x) for the following L_out frames; radar echoes use one measurement channel, C = 1.H and W denote frame spatial resolution, and C denotes measurements per space-time coordinate.Source: §3.1 · p. 3

3.2. Preliminary:Diffusion

• DDPMs learn to reverse a Markov noising process that progressively corrupts data, with forward states sampled from a Gaussian conditioned on the clean sample.The forward distribution has mean sqrt(alpha_bar_t)y^0 and covariance (1-alpha_bar_t)I.Source: §3.2 · p. 3
• The reverse process uses parameterized Gaussian transitions, with a trainable denoising function estimating the noise conditioned on the current diffusion step.The reverse transition aims to remove noise added by the forward process.Source: §3.2 · p. 3
• Training minimizes squared error between added noise and its estimate; generation starts from a Gaussian prior and iteratively applies the reverse update to obtain a target-distribution sample.The passage describes repeatedly applying the reverse step to recover the sample.Source: §3.2 · p. 3

4. Overall Framework

• DiffCast models precipitation as a global motion trend plus a local stochastic residual, combining a deterministic forecast with diffusion-predicted residuals.The residual is added to the deterministic prediction to form the final forecast.Source: §4 · p. 3

4.1. Deterministic Predictive Backbone

• The deterministic backbone can use recurrent or recurrent-free predictors to estimate the global motion trend, which also informs residual modeling.The backbone may use its native pixel-wise loss; its forecast can still be blurry and underestimate high-value echoes.Source: §4.1 · pp. 3–4

4.2. Stochastic Residual Prediction

• The diffusion component models residuals between ground-truth frames and the backbone forecast, predicting their evolution autoregressively from previous residuals.At inference, residuals are generated by iteratively denoising Gaussian noise; the framework combines denoising and deterministic losses for end-to-end training.Source: §4.2 · p. 4

4.3. Global Temporal UNet (GTUNet)

• GTUnet conditions residual prediction on global motion features, previous residual segments, segment position, and inter-frame temporal dependencies.Its GlobalNet derives motion information from the deterministic forecast, while temporal attention models dependencies within a segment.Source: §4.3 · p. 5

4.4. Training and Inference

• Training jointly optimizes the backbone and diffusion model; inference denoises residual segments autoregressively and adds them to the backbone forecast.Training computes residuals from ground truth, whereas inference samples and denoises segments before forming the final prediction.Source: §4.4 · p. 5

5.1. Experimental Setting

• The evaluation uses four radar datasets, with 20-frame forecasts from five input frames except CIKM, which uses 10 forecast frames.All datasets retain their original temporal resolution and are downscaled to 128 × 128.Source: §5.1 · p. 6

5.2. Experimental Results

• DiffCast improves backbone performance by 2%–20% on CSI and HSS, improves LPIPS by 16.9%–66.7%, and outperforms the GAN and diffusion baselines MCVD, PreDiff, and STRPM.The reported accuracy gains are larger for pooling CSI; the framework also improves SSIM.Source: §5.2 · p. 7
• As lead time increases, performance declines for all methods, but methods equipped with DiffCast remain better than their counterparts without it.Figure 4 plots CSI, HSS, and LPIPS for SimVP, Earthformer, and PhyDnet with and without the framework.Source: §5.2 · p. 7, Figure 4 · p. 7
• On a SEVIR event, DiffCast SimVP combines realistic detail with accurate placement and predicts two closing yellow rain belts at 100 minutes, unlike the comparison methods.SimVP is blurry, while MCVD, PreDiff, and STRPM produce detailed forecasts with less accurate positions.Source: §5.2 · p. 7, Figure 5 · p. 7

5.3. Analysis and Discussions

• Without deterministic loss (α = 1), predictions retain realistic detail but lose accuracy, whereas using both losses lets μ capture the global trend and residuals supply detail.The deterministic loss also improves convergence speed and quality; without it, the deterministic loss does not decrease.Source: §5.3 · p. 8, Figure 6 · p. 8
• End-to-end training outperforms the two-stage strategy, in which a pretrained deterministic backbone is frozen and diffusion acts as a corrector.The denoising loss also sends gradients to update the backbone during end-to-end training.Source: §5.3 · p. 8, Table 3 · p. 8
• Removing GlobalNet degrades performance, supporting its effectiveness in the framework.The ablation is on the SEVIR dataset.Source: §5.3 · p. 8, Table 4 · p. 8

• DiffCast decomposes precipitation-system evolution into a global deterministic trend and a local stochastic component using residual diffusion.Source: §6 · p. 8
• The framework is designed for precipitation nowcasting.Source: §6 · p. 8
• Experiments on four real-world datasets verify the framework’s effectiveness.Source: §6 · p. 8

SEVIR

• SEVIR forecasts 20 future VIL frames (100 minutes) from five observed frames, using 20,393 events and six thresholds for CSI and HSS.Events cover 384 km × 384 km patches sampled every five minutes; splits use time points January 1 and June 1, 2019.Source: §7 · p. 11

MeteoNet and Shanghai Radar

• MeteoNet covers northwestern France from 2016–2018, while Shanghai Radar records roughly six-minute scans over Pudong from 2015–2018; both use 25-frame event sequences.MeteoNet events are filtered with a stride-20 sliding window; Shanghai events are also filtered, and both datasets use dataset-specific CSI and HSS thresholds.Source: §7 · p. 11

CIKM

• CIKM samples 15 historical radar maps at six-minute intervals over Guangdong, pads maps to 128 × 128, and evaluates CSI and HSS at four thresholds.The samples cover 101 km × 101 km, and pixel values are converted to reflectivity in [0,76] dBZ.Source: §7 · p. 11

Forecast horizons

• Except for CIKM, the tasks forecast 20 frames, a longer horizon than the commonly used 10-frame, one-hour setup.The paper notes that predicting fewer frames at 12-minute intervals simplifies precipitation dynamics and reduces temporal resolution.Source: §7 · p. 11

Weather event filtering

• Weather-event filtering selects sequences around precipitation activity using pixel-threshold checks on frames and on the event-level mean.The algorithm considers five input frames and 20 output frames; accepted events advance the scan by the output length.Source: §7 · p. 11

Training procedure

• Training builds residual sequences from paired input and target sequences, groups them into segments, and combines denoising and deterministic losses in the final objective.The algorithm samples a diffusion step and updates the predictor, diffusion model, and associated parameters together.Source: §8 · p. 12

Training and inference

• Training jointly optimizes the base predictor and residual diffusion model; inference predicts a global trend, generates residual segments autoregressively, and combines both components.The procedures are summarized in Algorithms 2 and 3.Source: §8 · p. 12

Inference procedure

• Inference starts from a basic prediction and global hidden state, then applies reverse diffusion to generate residual segments before forming the target frames.Residual segments are generated in a loop until the forecast is complete.Source: §8 · p. 12

Implementation and cost

• GTUNet uses a hierarchical UNet with temporal attention; for 30K iterations, DiffCast MAU grows from 20.13MB to 66.40 MB and training rises from 18 hours at batch size 12 to 23 hours at batch size 6.DDIM and DPM-solver can expedite forecasts; producing 20-frame forecasts takes 17 seconds.Source: §8 · p. 12

Stochastic loss

• Stochastic loss positively optimizes most deterministic backbones, with Table 6 comparing backbone outputs under different optimization strategies on SEVIR.The framework is trained end-to-end, allowing stochastic diffusion denoising loss to optimize the deterministic backbone.Source: §9 · pp. 12–13, Table 6 · p. 13

High-value echoes

• With stochastic modeling, DiffCast SimVP maintains and predicts high-value echoes more accurately; its deterministic intermediate output also alleviates echo-value fading.Figure 8 presents CSI by threshold, and Figure 9 shows qualitative results without stochastic loss.Source: §9 · p. 13, Figure 8 · p. 13, Figure 9 · p. 13

Complexity and hyperparameter K

• K=5 provides the best reported tradeoff among tested segment lengths, while model size remains unchanged across K and memory use exceeds that of the base predictor.The comparison uses batch size 4 and 30K training iterations; the authors describe the added memory requirement as acceptable for their application.Source: §9 · p. 13, Table 7 · p. 13

Cross-dataset comparisons

• Across SEVIR, Shanghai Radar, MeteoNet, and CIKM examples, deterministic backbones become blurry after 60 minutes, while DiffCast-equipped predictions are consistently enhanced.The cited examples are Figures 10–13.Source: §10 · p. 13, Figure 10 · p. 14, Figure 11 · p. 15

Echo and detail preservation

• DiffCast-equipped forecasts are reported as no longer blurry, with high-value echoes and details preserved across the illustrated datasets.The comparison covers all backbones shown in the four dataset examples.Source: §10 · p. 13, Figure 10 · p. 14, Figure 11 · p. 15

Overall qualitative findings

• The paper reports consistent qualitative enhancement when deterministic backbones are equipped with DiffCast across the four illustrated datasets.The examples are presented as evidence of the framework's effectiveness.Source: §10 · p. 13, Figure 10 · p. 14, Figure 11 · p. 15