Observation-Free Representation Learning for Aerodrome Forecasting: The STATTCast and STATTTAF Pipeline

Short-term operational aviation weather guidance relies heavily on traditional linear statistical post-processing systems, such as the Localized Aviation MOS Program (LAMP; Ghirardelli and Glahn, 2010). While these systems achieve state-of-the-art (SOTA) skill in the 1-to 4-hour window by anchoring forecasts to real-time surface METAR observations (WMO, 2022), their skill decays rapidly beyond +6 h as boundary-layer persistence fades. Conversely, translating raw convection-allowing numerical weather prediction (NWP) columns directly into surface aviation hazards via machine learning is bottlenecked by sub-grid horizontal phase shifts and vertical “knife-edge” inversion noise. To resolve this dilemma, we introduce a two-stage, observation-free deep learning pipeline modeled after the cognitive synthesis of human operational meteorologists, in which both stages are conditioned on pre-trained “local-effects” geophysical representations. First, STATTCast (Short-Term Atmospheric Temporal Transformer foreCast) is a hierarchical 4D spatiotemporal transformer conditioned on a pre-trained hybrid convolutional-transformer coastal “marineness” autoencoder and spectral terrain encodings. Utilizing lossless volumetric space-todepth folding, disposable register tokens, and a neuro-inspired “Dream Cycle” latent manifold mixup regularizer, STATTCast ingests 82.5 km × 82.5 km mesoscale volumes from the 2.5 km Canadian High-Resolution Deterministic Prediction System (HRDPS; Milbrandt et al., 2016) and projects them onto a denoised, 38-level ERA5-aligned (Hersbach et al., 2020) vertical profile manifold. Second, STATTTAF, a 2D Anisotropic Multi-Gate Mixture-of-Experts (MMoE) transformer that expands the local-effects conditioning to pair the marineness autoencoder with localized Daubechies-4 (db4) wavelet topography (Daubechies, 1992) and continental wavelet mean sea-level pressure (MSLP) encodings, translates instantaneous vertical columns into a 27-dimensional vector of ceiling, visibility, present weather, and surface wind guidance. Pre-training STATTTAF across a multi-decadal reanalysis archive is enabled by a novel surface-isolated 2D-Var background error covariance (B-matrix) augmentation scheme. Evaluated out-of-sample against n = 1, 070 operational radiosonde launches across 12 Canadian upper-air stations, STATTCast reduces free-tropospheric temperature and dew point root-mean-square error (RMSE) by 56.9% (to 0.91 ◦C) and 34.7% (to 3.97 ◦C), respectively, and cuts surface wind speed RMSE by 52.2% (to 2.04 m s−1) relative to raw HRDPS. In a synchronized end-to-end ablation across N = 12, 591 forecast hours at 10 major Canadian hub aerodromes, projecting NWP output through the STATTCast manifold improves 4-class Heidke Skill Score (HSS; Heidke, 1926) by +29.2% and Low Instrument Flight Rules (LIFR) recall by +48.6% compared to an identical network trained directly on native HRDPS columns. Coupled with a heuristic cascading boundary-layer saturation event post-processor derived from operational forecasting rules (Filter-50), the observation-free STATTCast → STATTTAF pipeline outperforms operational HRRR–LAMP benchmarks (Glahn et al., 2017) across 7 of 8 seasonal ceiling and visibility Threat Score categories over the F + 06 to F + 17 flight-planning horizon while achieving SOTA surface wind accuracy (1.34 m s−1 MAE; 1.74 m s−1 RMSE).

Authors

Institutions

Publication Details

Published
2026-11-20
DOI
https://doi.org/10.22541/essoar.15012534/v1
Primary Topic
Meteorological Phenomena and Simulations
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Observation-Free Representation Learning for Aerodrome Forecasting: The STATTCast and STATTTAF Pipeline

Quenton Hurdal
Meteorological Phenomena and Simulations
preprint

Observation-Free Representation Learning for Aerodrome Forecasting: The STATTCast and STATTTAF Pipeline

Quenton Hurdal
preprint en

Abstract

Short-term operational aviation weather guidance relies heavily on traditional linear statistical post-processing systems, such as the Localized Aviation MOS Program (LAMP; Ghirardelli and Glahn, 2010). While these systems achieve state-of-the-art (SOTA) skill in the 1-to 4-hour window by anchoring forecasts to real-time surface METAR observations (WMO, 2022), their skill decays rapidly beyond +6 h as boundary-layer persistence fades. Conversely, translating raw convection-allowing numerical weather prediction (NWP) columns directly into surface aviation hazards via machine learning is bottlenecked by sub-grid horizontal phase shifts and vertical “knife-edge” inversion noise. To resolve this dilemma, we introduce a two-stage, observation-free deep learning pipeline modeled after the cognitive synthesis of human operational meteorologists, in which both stages are conditioned on pre-trained “local-effects” geophysical representations. First, STATTCast (Short-Term Atmospheric Temporal Transformer foreCast) is a hierarchical 4D spatiotemporal transformer conditioned on a pre-trained hybrid convolutional-transformer coastal “marineness” autoencoder and spectral terrain encodings. Utilizing lossless volumetric space-todepth folding, disposable register tokens, and a neuro-inspired “Dream Cycle” latent manifold mixup regularizer, STATTCast ingests 82.5 km × 82.5 km mesoscale volumes from the 2.5 km Canadian High-Resolution Deterministic Prediction System (HRDPS; Milbrandt et al., 2016) and projects them onto a denoised, 38-level ERA5-aligned (Hersbach et al., 2020) vertical profile manifold. Second, STATTTAF, a 2D Anisotropic Multi-Gate Mixture-of-Experts (MMoE) transformer that expands the local-effects conditioning to pair the marineness autoencoder with localized Daubechies-4 (db4) wavelet topography (Daubechies, 1992) and continental wavelet mean sea-level pressure (MSLP) encodings, translates instantaneous vertical columns into a 27-dimensional vector of ceiling, visibility, present weather, and surface wind guidance. Pre-training STATTTAF across a multi-decadal reanalysis archive is enabled by a novel surface-isolated 2D-Var background error covariance (B-matrix) augmentation scheme. Evaluated out-of-sample against n = 1, 070 operational radiosonde launches across 12 Canadian upper-air stations, STATTCast reduces free-tropospheric temperature and dew point root-mean-square error (RMSE) by 56.9% (to 0.91 ◦C) and 34.7% (to 3.97 ◦C), respectively, and cuts surface wind speed RMSE by 52.2% (to 2.04 m s−1) relative to raw HRDPS. In a synchronized end-to-end ablation across N = 12, 591 forecast hours at 10 major Canadian hub aerodromes, projecting NWP output through the STATTCast manifold improves 4-class Heidke Skill Score (HSS; Heidke, 1926) by +29.2% and Low Instrument Flight Rules (LIFR) recall by +48.6% compared to an identical network trained directly on native HRDPS columns. Coupled with a heuristic cascading boundary-layer saturation event post-processor derived from operational forecasting rules (Filter-50), the observation-free STATTCast → STATTTAF pipeline outperforms operational HRRR–LAMP benchmarks (Glahn et al., 2017) across 7 of 8 seasonal ceiling and visibility Threat Score categories over the F + 06 to F + 17 flight-planning horizon while achieving SOTA surface wind accuracy (1.34 m s−1 MAE; 1.74 m s−1 RMSE).

Environment and Climate Change Canada (CA)
Meteorological Phenomena and Simulations
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon