Singh et al. (2026) START: A Hybrid Spatio-Temporal Attention ResNet Transformer for Explainable Multivariable Meteorological Bias-correction

Identification

• Journal: Earth Systems and Environment
• Year: 2026
• Date: 2026-03-26
• Authors: Deveshwar Singh, Yunsoo Choi, Rijul Dimri
• DOI: 10.1007/s41748-026-01132-4

Research Groups

• Department of Earth and Atmospheric Sciences, College of Natural Science and Mathematics, University of Houston, TX, USA

Short Summary

This study introduces START, a hybrid deep learning framework for multivariable meteorological bias correction over the contiguous United States, integrating heterogeneous data streams to achieve substantial improvements in forecast accuracy and provide explainable, calibrated uncertainty estimates.

Objective

• To develop and evaluate START, a hybrid deep learning framework, for multivariable meteorological bias correction over the contiguous United States, aiming to enhance forecast accuracy, preserve inter-variable dependencies, and provide explainable, calibrated uncertainty estimates.

Study Configuration

• Spatial Scale: Contiguous United States (CONUS), defined by latitudes 24.52°N to 49.38°N and longitudes 66.95°W to 124.77°W. Spatial resolution: approximately 0.25° × 0.25° (25 km × 25 km).
• Temporal Scale: Study period: January 1, 2019, to December 31, 2022. Training data: 2019-2021. Evaluation data: 2022. Temporal resolution: hourly data.

Methodology and Data

• Models used: START (Spatio-Temporal Attention ResNet Transformer), a hybrid deep learning architecture combining ResNet blocks, attention pooling, and transformer-based fusion.
• Data sources:
  • NASA GEOS-CF v1.0: Numerical weather prediction model output, providing 23 forecast variables (reduced to 12 after feature selection).
  • ERA5 reanalysis: Reference observational dataset, providing 5 previous-day observational variables (e.g., T2Mobs, PSobs, RHobs, WSPDobs, WDIRobs).
  • Positional encoding features: 16 features derived from geographical coordinates (latitude, longitude).
  • Temporal encoding features: 6 features derived from time (hour of day, day of year, day of week).
  • Total input features: 17 meteorological variables + 16 positional encoding features + 6 temporal encoding features = 39 features.

Main Results

• The START model achieved substantial Index of Agreement (IOA) improvements over the NASA GEOS baseline on unseen 2022 test data:
  • Temperature: IOA = 0.99 (1.1% improvement)
  • Pressure: IOA = 0.98 (4.5% improvement)
  • Wind speed: IOA = 0.92 (7.7% improvement)
  • Wind direction: circular IOA = 0.95 (0.2% improvement)
  • Relative humidity: IOA = 0.93 (7.0% improvement)
• SHAP analysis revealed that meteorological inputs predominantly contributed (75–85%) to predictions across all variables.
• Spatial encodings significantly influenced pressure (~30%) and wind direction (~20%) predictions, reflecting the importance of geographic context.
• Temporal features provided seasonal and diurnal context, contributing 5–15% across variables.
• Monte Carlo Dropout (MCD) uncertainty quantification yielded epistemic uncertainty estimates that positively correlated with prediction errors (r = 0.27–0.49).
• Post-hoc second-order polynomial calibration effectively rectified raw uncertainty miscalibration, achieving near-perfect spread-skill alignment (SSREL < 0.04) across all variables, enabling reliable probabilistic forecasting.
• A sequential transfer learning strategy reduced validation loss by 22% and enhanced training stability.

Contributions

• Introduced START, a novel hybrid deep learning architecture for multivariable meteorological bias correction, integrating heterogeneous data streams (meteorological, spatial positional encodings, temporal features) through ResNet blocks, attention pooling, and transformer-based fusion.
• Demonstrated superior performance over the NASA GEOS operational baseline across five key meteorological variables (temperature, pressure, wind speed, wind direction, relative humidity) over the contiguous United States, with physically constrained output scaling and preservation of inter-variable dependencies.
• Provided explainable AI insights via SHAP analysis, elucidating the differential contributions of various input modalities to predictions and validating architectural design choices.
• Implemented and calibrated uncertainty quantification using Monte Carlo Dropout and second-order polynomial calibration, enabling reliable probabilistic forecasts with near-perfect spread-skill alignment.
• Employed a sequential transfer learning strategy for stable training and enhanced generalizability across different years.

Funding

• High Priority Area Research Grant of the University of Houston.
• University of Houston Hewlett Packard Enterprise Data Science Institute for providing HPC resources.

Citation

@article{Singh2026START,
  author = {Singh, Deveshwar and Choi, Yunsoo and Dimri, Rijul},
  title = {START: A Hybrid Spatio-Temporal Attention ResNet Transformer for Explainable Multivariable Meteorological Bias-correction},
  journal = {Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s41748-026-01132-4},
  url = {https://doi.org/10.1007/s41748-026-01132-4}
}