PaperlayerView exact revision on arXiv

Source-linked AI summary

Methodological Changes to the Attention ResUNet Hourly Precipitation Postprocessor

Thomas M. Hamill

arXiv:2609.38609v1physics.ao-phcs.LG

TL;DR

The original feature set did not represent possible diurnal dependence in convective precipitation errors. The updated postprocessor adds explicit inputs and uses seasonal FiLM conditioning to train four models across the full lead range. Current training retains general reliability and improves forecast skill against a climatological reference.

  • Problem

    The original feature set did not represent possible diurnal dependence of convective precipitation errors.

    Source: §2.1 · p. 3

  • Method

    The updated method adds local-solar-hour and climatology inputs and uses FiLM conditioning on calendar time and lead time to train one model per season across leads.

    Source: §2.1 · p. 3, §2.2 · p. 3, §2.3 · p. 4

  • Results

    Current training retained general reliability and improved climatological-reference forecast skill; over roughest terrain at 5 mm, BSS rose from 0.121 to 0.176 at 6 h and from 0.078 to 0.116 at 48 h.

    Source: §3.2 · p. 5, §4 · pp. 8–9

  • Takeaways & Limitations

    Four seasonal checkpoints now cover the continuous +3 to +72 h lead range instead of the original 192 month-and-lead checkpoints covering +3 to +48 h.

    Source: §2.3 · p. 4

    At the highest forecast probabilities, the 5 mm curves show sampling irregularity where usage frequency falls below 10^-3.

    Source: §3.2 · p. 7

Abstract

from arXiv · show

This note is a technical companion to a previously published preprint describing an Attention Residual U-Net that postprocesses deterministic forecasts from The Weather Company's Global and Regional Atmospheric Forecast (GRAF) model into probabilistic hourly precipitation forecasts. It documents what has changed in that method since publication. Feature-wise Linear Modulation conditioning on calendar season and forecast lead time is used to produce a single trained model for each season, replacing 192 separately trained per-month, per-lead checkpoints. Lead time is extended from 48 to 72 h. Two new input channels are used, per-pixel local solar hour and a static, monthly-varying precipitation climatology. During verification, the climatological reference against which the Brier Skill Score is computed now has an added diurnal dimension, on top of the monthly resolution it already had. Brier Skill Score and reliability are compared between the new vs. the previous training. Forecasts generated with the new training show a modest, consistent improvement of the current training over the original.

1 Introduction and Scope

This note documents changes to the Attention ResUNet precipitation postprocessor since its original preprint, covering only the most recent configuration. Its sections follow the original manuscript’s structure for side-by-side reading.

  • The note records changes to the previously described Attention ResUNet probabilistic precipitation postprocessor.Source: §1 · p. 1
  • It documents only the most recent configuration, although other algorithmic improvements were tested after the first version.Source: §1 · p. 1
  • The note separates modeling-method changes from verification changes and follows the original manuscript’s section structure.Source: §1 · p. 1

2 Methodological Changes

The update retains the original model’s core architecture while changing its inputs, conditioning, training-data pipeline, and inference selection. Seasonal FiLM conditioning supports four models spanning the continuous +3 to +72 h range.

2 Methodological Changes

  • The Attention ResUNet architecture, predictive distribution, loss, climatological bias initialization, and fully convolutional whole-domain inference remain unchanged.Source: §2 · p. 1

2.1 Input Features and Training Data

  • Local solar hour, computed from UTC valid hour and each pixel’s longitude, adds an explicit input for possible diurnal variation in convective precipitation errors.Source: §2.1 · pp. 2–3
  • A static, per-pixel monthly precipitation climatology adds an explicit prior on locations and months that are climatologically wetter or drier.Source: §2.1 · p. 3
  • Season-specific zarr-backed archives pool patches across available years and lead times, while preserving wet-day-weighted sampling and assigning contiguous seven-day blocks to training or holdout.Source: §2.1 · p. 3

2.2 Network Conditioning (FiLM)

  • FiLM maps day-of-year sine/cosine and normalized lead time to channel-wise scale and shift values at encoder, bridge, and decoder stages.Source: §2.2 · p. 3

2.3 Training Regime

  • Training changes from up to 192 month-and-lead checkpoints covering +3 to +48 h to four seasonal checkpoints covering +3 to +72 h continuously; core optimization settings remain unchanged.Source: §2.3 · p. 4

2.4 Inference Procedure

  • Inference selects a seasonal checkpoint by calendar month and supplies the requested lead and day of year directly, replacing nearest-lead selection among up to 192 checkpoints.Source: §2.4 · p. 4

3 Verification and Evaluation Methodology Changes

Verification retains the established metrics and terrain stratification but adds a diurnal dimension to the Brier Skill Score climatology. On the shared evaluation sample, the current training modestly improves BSS over the original, with reliability and case-study comparisons also reported.

3.1 Climatological Reference Refinement

  • The BSS climatology adds all 24 UTC hours to monthly resolution because monthly means cannot resolve convection's diurnal cycle at 1–3 h forecast leads.The interpolation was also changed from bilinear interpolation to Delaunay triangulation with linear barycentric interpolation.Source: §3.1 · pp. 4–5

3.2 Comparing Forecast Skill: Original vs. Current Training

  • Both trainings use the same March, June, September, and December 2025 six-hourly sample, while comparison with the original Gaussian-smoothed, thresholded GRAF reference is omitted.The sample contains 447 initialization times per lead with complete GRAF, MRMS, and probability data.Source: §3.2 · p. 5
  • Figure 1 uncertainty bands are 90% Hodges-Lehmann intervals from per-cycle BSS values, recentered on each plotted score rather than the median-targeting location estimate.Each 6-hourly verification cycle pools grid points into one score.Source: §3.2 · p. 5, Figure 1 · p. 6
  • Current training matches or exceeds original-training BSS at every shown lead, threshold, and terrain stratum, with the largest gains at 5 mm over rough terrain.At 6 h, BSS rises from 0.121 to 0.176; at 48 h, it rises from 0.078 to 0.116.Source: §3.2 · p. 5, Figure 1 · p. 6
  • At +12 h and +48 h, both trainings are well calibrated, while current-training BSS is higher at every displayed threshold; high-probability 5 mm curves remain irregular below 10^-3 usage frequency.The reliability diagrams cover exceedances of 0.25, 1.0, and 5.0 mm.Source: §3.2 · pp. 6–7, Figure 2 · p. 7, Figure 3 · p. 7
  • In the Washington atmospheric-river case, both postprocessed fields place high probabilities along the Olympic Peninsula and Cascades rather than the raw forecast's offshore-centered maximum.The clearest difference between the two postprocessed panels is the areal coverage of high probability.Source: §3.2 · pp. 7–8, Figure 4 · p. 8

4 Summary

The statistical postprocessor remains unchanged, while its conditioning, training, and verification have been updated; newly trained forecasts retain general reliability and improve skill against a climatological reference.

  • The core statistical method remains unchanged; updates concern conditioning and training across lead times and seasons, and how the model is verified.Source: §4 · p. 8
  • FiLM conditioning lets one model per season cover +3 to +72 h, replacing up to 192 per-month, per-lead checkpoints covering only +3 to +48 h.Single-pass whole-domain inference was preserved by making spatially varying solar hour a per-pixel input rather than part of the FiLM vector.Source: §4 · p. 8
  • Verification now uses a canonical diurnally resolved BSS reference and supports cached-score reliability comparisons across successive model retrains.The reference change improves provenance and resolution, rather than making a large numerical correction.Source: §4 · p. 8
  • Newly trained forecasts retained general reliability while improving skill against a climatological reference.Source: §4 · pp. 8–9
Loading 2609.38609v1…