Scholarlink

<p>Accurate meteorological prediction is critical for applications but faces systematic biases in numerical weather prediction models. This study presents the Spatial Temporal Attention ResNet Transformer (START), a deep learning framework for multivariable meteorological bias correction over the contiguous United States. START integrates three heterogeneous data streams through a hybrid architecture combining ResNet-based feature extraction, attention pooling, and transformer-based fusion to capture complex spatiotemporal dependencies. A two-stage feature selection process using hierarchical clustering and variance inflation factor analysis reduced 28 candidate predictors to 17 meteorological variables, supplemented by 16 positional encoding features and 6 temporal encoding features. The model was trained on hourly data from 2019 to 2021 using NASA GEOS-CF and ERA5 datasets, employing sequential transfer learning to enhance stability, achieving 22% reduction in validation loss. Evaluation on unseen 2022 test data demonstrated substantial Index of Agreement (IOA) improvements over NASA GEOS baseline: 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), and relative humidity (IOA = 0.93, 7.0% improvement). SHAP analysis revealed meteorological inputs contributed predominantly (75–85%) to predictions, with spatial encodings significantly influencing pressure (~ 30%) and wind direction (~ 20%), while temporal features provided seasonal context (5–15%). Monte Carlo Dropout uncertainty quantification yielded epistemic uncertainty estimates correlated with prediction errors (<i>r</i> = 0.27–0.49). Second-order polynomial calibration achieved near-perfect spread-skill alignment (SSREL &lt; 0.04), enabling reliable probabilistic forecasting. START advances meteorological bias correction through physically constrained scaling, inter-variable dependency preservation, and spatial-temporal embeddings, demonstrating deep learning’s potential to enhance numerical weather prediction systems.</p> Graphical Abstract <p></p> <p>This graphical abstract illustrates the START (Spatio-Temporal Attention ResNet Transformer) model, an advanced deep learning framework designed for bias-correcting meteorological predictions across the Continental United States. The workflow begins with the integration of NASA GEOS and ERA5 reanalysis datasets, collectively providing 28 meteorological variables including precipitation, temperature, pressure, wind parameters, and various atmospheric conditions. Through hierarchical clustering combined with Variance Inflation Factor analysis, the initial variable set undergoes systematic dimensionality reduction to 17 essential predictors, eliminating redundancy while preserving critical meteorological information as demonstrated by the dendrogram visualization. The model architecture features a sophisticated combination of multi-stream transformer design where three parallel ResNet blocks process distinct information streams: selected meteorological variables, geo-encoded spatial coordinates with positional encoding, and temporal patterns via time stamp encoding. These streams converge through transformer blocks and an attention pooling mechanism that synthesizes five fundamental components namely, temperature, pressure, wind speed, wind direction, and relative humidity, capturing complex spatio-temporal dependencies inherent in atmospheric processes. The framework produces three key outputs demonstrating its effectiveness: bias-corrected predictions for five meteorological variables across CONUS, SHAP analysis providing transparent stream and feature importance quantification that reveals which variables drive predictions and enhances scientific interpretability, and comprehensive uncertainty quantification through validation plots comparing model outputs against observed data with confidence intervals. By combining transformer architecture with explainable AI techniques and rigorous uncertainty assessment, START successfully delivers accurate, interpretable, and reliable meteorological bias corrections suitable for operational weather forecasting, climate modeling, and decision-making processes requiring trustworthy atmospheric predictions across continental spatial domains while maintaining the scientific transparency essential for advancing atmospheric science applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

START: A Hybrid Spatio-Temporal Attention ResNet Transformer for Explainable Multivariable Meteorological Bias-correction

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Deveshwar Singh,
  • Yunsoo Choi,
  • Rijul Dimri

摘要

Accurate meteorological prediction is critical for applications but faces systematic biases in numerical weather prediction models. This study presents the Spatial Temporal Attention ResNet Transformer (START), a deep learning framework for multivariable meteorological bias correction over the contiguous United States. START integrates three heterogeneous data streams through a hybrid architecture combining ResNet-based feature extraction, attention pooling, and transformer-based fusion to capture complex spatiotemporal dependencies. A two-stage feature selection process using hierarchical clustering and variance inflation factor analysis reduced 28 candidate predictors to 17 meteorological variables, supplemented by 16 positional encoding features and 6 temporal encoding features. The model was trained on hourly data from 2019 to 2021 using NASA GEOS-CF and ERA5 datasets, employing sequential transfer learning to enhance stability, achieving 22% reduction in validation loss. Evaluation on unseen 2022 test data demonstrated substantial Index of Agreement (IOA) improvements over NASA GEOS baseline: 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), and relative humidity (IOA = 0.93, 7.0% improvement). SHAP analysis revealed meteorological inputs contributed predominantly (75–85%) to predictions, with spatial encodings significantly influencing pressure (~ 30%) and wind direction (~ 20%), while temporal features provided seasonal context (5–15%). Monte Carlo Dropout uncertainty quantification yielded epistemic uncertainty estimates correlated with prediction errors (r = 0.27–0.49). Second-order polynomial calibration achieved near-perfect spread-skill alignment (SSREL < 0.04), enabling reliable probabilistic forecasting. START advances meteorological bias correction through physically constrained scaling, inter-variable dependency preservation, and spatial-temporal embeddings, demonstrating deep learning’s potential to enhance numerical weather prediction systems.

Graphical Abstract

This graphical abstract illustrates the START (Spatio-Temporal Attention ResNet Transformer) model, an advanced deep learning framework designed for bias-correcting meteorological predictions across the Continental United States. The workflow begins with the integration of NASA GEOS and ERA5 reanalysis datasets, collectively providing 28 meteorological variables including precipitation, temperature, pressure, wind parameters, and various atmospheric conditions. Through hierarchical clustering combined with Variance Inflation Factor analysis, the initial variable set undergoes systematic dimensionality reduction to 17 essential predictors, eliminating redundancy while preserving critical meteorological information as demonstrated by the dendrogram visualization. The model architecture features a sophisticated combination of multi-stream transformer design where three parallel ResNet blocks process distinct information streams: selected meteorological variables, geo-encoded spatial coordinates with positional encoding, and temporal patterns via time stamp encoding. These streams converge through transformer blocks and an attention pooling mechanism that synthesizes five fundamental components namely, temperature, pressure, wind speed, wind direction, and relative humidity, capturing complex spatio-temporal dependencies inherent in atmospheric processes. The framework produces three key outputs demonstrating its effectiveness: bias-corrected predictions for five meteorological variables across CONUS, SHAP analysis providing transparent stream and feature importance quantification that reveals which variables drive predictions and enhances scientific interpretability, and comprehensive uncertainty quantification through validation plots comparing model outputs against observed data with confidence intervals. By combining transformer architecture with explainable AI techniques and rigorous uncertainty assessment, START successfully delivers accurate, interpretable, and reliable meteorological bias corrections suitable for operational weather forecasting, climate modeling, and decision-making processes requiring trustworthy atmospheric predictions across continental spatial domains while maintaining the scientific transparency essential for advancing atmospheric science applications.