Statistical treatment of convolutional neural network super-resolution of inland surface wind for subgrid-scale variability quantification

Daniel J. Getter, Julie Bessac, Johann Rudi, Feng YanarXiv 2022 · DOI · arXiv

TL;DR

Researchers have developed a Convolutional Neural Network (CNN) framework to "super-resolve" coarse climate model wind data, downscaling it from 100 km resolution to a sharp 3 km. By incorporating fine-scale topography and probabilistic modeling, the system doesn't just predict average wind speeds—it captures the complex subgrid variability and extreme wind events essential for wind energy and hazard assessment.

What Problem Does This Paper Address?

Global Climate Models (GCMs) are the workhorses of climate science, but they have a "blurry" vision problem. Due to massive computational costs, they typically produce data at resolutions of 50–100 km. However, the real world happens at much smaller scales: wind gusts are influenced by local hills, valleys, and urban heat islands.

For industries like wind energy or disaster management (predicting storm surges and flooding), a 100 km average is nearly useless. We need to know what’s happening at the 3 km scale. While running high-resolution physical simulations is one solution, it is prohibitively expensive. This paper explores whether deep learning can bridge this "resolution gap" by learning the relationship between coarse data and fine-scale reality.

Methods

The authors propose SR-CNN (Super-Resolution Convolutional Neural Network), testing several configurations to see which inputs matter most:

  1. 1-Channel: Uses only coarse wind speed.
  2. 2-Channel: Adds high-resolution topography (terrain height) as a second input.
  3. 3-Channel: Adds a temporal predictor (diurnal cycle) to capture how wind changes throughout the day.

Beyond standard deterministic models (which predict a single value), they introduced a Probabilistic Gumbel Loss model. Instead of predicting one wind speed, this version predicts the parameters of a Gumbel distribution. This allows the model to account for the "intrinsic stochasticity" (randomness) of wind, providing a range of possible values and better capturing the "tails" of the distribution—where the dangerous, high-speed extreme winds live.

Datasets

The study uses data from SCREAM (Simple Cloud-Resolving E3SM Atmosphere Model), a state-of-the-art global model that provides 3.25 km resolution data.

  • Training/Validation: 40 days of 15-minute interval data over the Southwestern United States.
  • Generalizability Testing: To see if the model learned physics or just memorized a map, they tested it on two completely unseen regions: the Eastern Sahara and Eastern Asia.

Key Insights and Ideas

  • Topography is King: The 2-channel model (wind + terrain) significantly outperformed the wind-only model. This confirms that local geography is a primary driver of wind variability that the CNN can effectively exploit.
  • The "Aha" on Extremes: Standard models trained on Mean Squared Error (MSE) tend to "average out" results, missing the highest and lowest winds. By using the Gumbel distribution, the researchers could sample realistic extremes that match the statistical profile of real-world wind.
  • The Diurnal Difficulty: Surprisingly, adding a simple timestamp (diurnal cycle) as a third channel actually reduced the model's ability to generalize to new regions. The authors suggest that wind's relationship with time is too complex for a simple periodic function and might require more physical inputs like solar irradiance.
  • Scale Matters: The model was tested at 8x, 16x, and 32x downscaling factors. While accuracy drops as the gap widens, the CNNs remained remarkably robust even at the 32x jump (100 km to 3 km).

Key Results

  • Superiority over Baselines: All CNN configurations crushed standard bicubic interpolation. In the U.S. test region, CNNs captured wind variability (measured by Jensen–Shannon distance) 3 to 5 times better than interpolation.
  • Low Error: Median Absolute Percentage Error (MAPE) remained low across all gaps. For the 25 km to 3 km jump, errors were often below 5%.
  • Extreme Recovery: The 2-channel models were particularly adept at capturing the 95th percentile winds, which are usually lost in coarse models.
  • Out-of-Sample Success: The models performed well in the Eastern Sahara, proving they had learned generalizable relationships between terrain and wind, though performance was lower in Eastern Asia due to its more complex, high-variability wind patterns.

Limitations and Future Work

The authors identify a few clear paths for improvement:

  • Temporal Granularity: The current diurnal encoding is too simple. Future versions might use solar radiation data to better inform the model about day/night atmospheric stability.
  • Spatial Correlation: While the Gumbel model predicts the right distribution at each point, it doesn't yet ensure that neighboring points are physically "linked" in a coherent spatial structure.
  • Unstructured Grids: Most climate models don't use nice, square pixels (like the ones used here). Adapting SR-CNN to work on the native, unstructured grids of climate models would make it much more practical for operational use.

Reference: Getter, D., Bessac, J., Rudi, J., & Feng, Y. (2023). Statistical treatment of convolutional neural network super-resolution of inland surface wind for subgrid-scale variability quantification. arXiv:2211.16708v3.

Summary updated

Abstract

Machine learning models have been employed to perform either physics-free data-driven or hybrid dynamical downscaling of climate data. Most of these implementations operate over relatively small downscaling factors because of the challenge of recovering fine-scale information from coarse data. This limits their compatibility with many global climate model outputs, often available between $\sim$50--100 km resolution, to scales of interest such as cloud resolving or urban scales. This study systematically examines the capability of convolutional neural networks (CNNs) to downscale surface wind speed data over land surface from different coarse resolutions (25 km, 48 km, and 100 km resolution) to 3 km. For each downscaling factor, we consider three CNN configurations that generate super-resolved predictions of fine-scale wind speed, which take between 1 to 3 input fields: coarse wind speed, fine-scale topography, and diurnal cycle. In addition to fine-scale wind speeds, probability density function parameters are generated, through which sample wind speeds can be generated accounting for the intrinsic stochasticity of wind speed. For generalizability assessment, CNN models are tested on regions with different topography and climate that are unseen during training. The evaluation of super-resolved predictions focuses on subgrid-scale variability and the recovery of extremes. Models with coarse wind and fine topography as inputs exhibit the best performance compared with other model configurations, operating across the same downscaling factor. Our diurnal cycle encoding results in lower out-of-sample generalizability compared with other input configurations.
BibTeX
@misc{getter2022statistical,
  author = {Daniel J. Getter and Julie Bessac and Johann Rudi and Feng Yan},
  title = {{Statistical treatment of convolutional neural network super-resolution of inland surface wind for subgrid-scale variability quantification}},
  year = {2022},
  eprint = {2211.16708},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.16708},
}
Figure