WeatherDiagFlow: Evidence-Grounded Radar Nowcasting with Diagnostic Flow Refinement

1. INTRODUCTION

• Radar forecasts must retain coherent motion and localized high-reflectivity structures across lead times, since mean absolute error alone can reward smoothing that removes rare strong echoes.Errors accumulate across successive forecast chunks, making late-lead evolution especially vulnerable to spatial smoothing and weak-event bias.Source: §1 · p. 1
• Operational nowcasting needs traceable evidence about forecast changes, locations, and confidence, with a boundary keeping future observations out of issued guidance.Numerical nowcasting, reporting, and verification remain largely separate, and free-form reports may contain claims unsupported by numerical forecasts.Source: §1 · p. 1
• WeatherDiagFlow predicts motion, growth and decay, heavy-echo risk, and uncertainty to condition forecasting, while separating forecast-time bulletins from post-event audits.The audit uses future radar truth only after it becomes available; textual agents package and check evidence without changing the numerical forecast.Source: §1 · p. 1, Figure 1 · p. 2
• Experiments on FJRADAR show competitive overall performance and improved strong-echo event skill with evidence-consistent bulletins and audits.The system's contributions include diagnostic-conditioned rolling flow refinement and frozen-scaffold residual calibration to preserve strong echoes.Source: §1 · p. 1

2.1. Coarse future prior

• Ten six-minute historical reflectivity frames feed a coarse forecaster that supplies a deterministic scaffold for broad displacement and coverage, leaving local evolution to later refinement.The system predicts thirty future frames over three hours; valid-range masking excludes invalid radar pixels from losses and metrics.Source: §2.1 · p. 2

2.2. Diagnostic field prediction

• At inference, the diagnostic encoder uses history and the coarse forecast to predict motion, growth or decay, heavy-echo risk, and uncertainty; ground-truth-derived targets supervise training only.Motion is constrained by brightness constancy and smoothness, and diagnostic targets are never forecast-time conditions.Source: §2.2 · p. 2

2.3. Rolling flow refinement

• The 30-frame horizon is refined in three one-hour chunks using rolling context and a pixel-space flow model, with a frozen scaffold and residual correction to limit feedback from earlier outputs.The residual path starts from low-amplitude noise, focusing correction on deviations not explained by the scaffold.Source: §2.3 · pp. 2–3
• Residual constraints target excessive echo-area growth, disappearance of persistent echoes, and inaccurate generated histories; positive and negative corrections are scaled separately and next-chunk context is scaffold-anchored.The anti-decay extension adds penalties for missed echoes, soft threshold skill, echo area and mass, and later forecast chunks; its variant is selected using validation data before independent testing.Source: §2.3 · p. 3

2.4. Forecast–bulletin–audit agents

• The agent layer creates a machine-readable evidence record and forecast-time bulletins from model-available quantities, then uses future radar truth for post-event audits without feeding text back to the forecaster.The products cover chunked evolution, attention regions, and evidence-bounded confidence; role-specific agents assess consistency, visual agreement, and bulletin quality, with future-truth leakage rejected.Source: §2.4 · p. 3

3.1. Dataset and protocol

• The event-focused FJRADAR evaluation uses a chronological 30,665/6,345/2,125 train/validation/test split and a common 2,125-case test manifest, scoring pooled valid pixels across 30 future frames.Cases have ten historical and thirty future frames at six-minute spacing; the subset retains cases with at least 5% valid future-frame pixels above 10 dBZ.Source: §3.1 · p. 3

3.2. Quantitative nowcasting results

• CSI30 rises from 0.0169 for matched rolling pMF to 0.1284 for WDF-RC and 0.1351 for WDF-RC-AD; WDF-RC-AD leads retained methods on SSIM, LPIPS, CSI20/30, POD20/30, and HSS20/30.CasCast remains strongest at CSI10 and competitive at CSI20.Source: §3.2 · p. 3, Table 1 · p. 4
• The diagnostic branch’s heavy-echo risk map reaches 0.2078 CSI at the 10-dBZ supervision threshold, and uncertainty correlates 0.3706 with realized absolute error.The authors note that motion is checked through brightness-constancy residual rather than ground-truth optical flow, and event-level calibration remains future work.Source: §3.2 · p. 3

3.3. Case-level and operational diagnosis

• Across six lead times over 0–3 h, WDF-RC-AD retains strong-echo structure while limiting late-lead area inflation; forecast products exclude ground truth, which is reserved for post-event audits.The 32-case task set scores report quality for evidence consistency, temporal coherence, spatial correctness, completeness, and operational usefulness.Source: §3.3 · pp. 3–4, Figure 3 · p. 4

3.4. Ablation study

• Residual calibration raises CSI30/POD30 from 0.0274/0.0298 to 0.1284/0.3005, and AD further reaches 0.1351/0.3384 with HSS30=0.2273 and FAR30=0.8164.The reported result supports anti-decay calibration as a threshold-skill adjustment rather than unconstrained intensity amplification, while preserving the evidence interface.Source: §3.4 · p. 4, Table 3 · p. 4

• WeatherDiagFlow formulates 0–3 hour radar prediction as an evidence-grounded forecast–bulletin–audit task.Source: §4 · p. 4
• Predicted motion, echo growth/decay, heavy-echo risk, and uncertainty condition rolling flow refinement, while WDF-RC adds frozen-scaffold residual calibration for stable long-lead prediction.Source: §4 · p. 4
• Future work will finalize FAR-constrained anti-decay selection and evaluate diagnostic calibration and bulletin usefulness across broader events.Source: §4 · p. 4

• The study uses radar reflectivity data.Source: §5 · p. 5
• Evaluation is automated and model-based.Source: §5 · p. 5
• No human participants or animals were involved, so ethical approval was not required.Source: §5 · p. 5