Generating High Resolution Climate Change Projections through Single Image Super-Resolution: An Abridged Version

Thomas Vandal, Evan Kodra, Sangram Ganguly, Andrew R. Michaelis, Ramakrishna R. Nemani, Auroop R. GangulyIJCAI 2018 · DOI

TL;DR

This paper introduces DeepSD, a novel deep learning framework that leverages single image super-resolution (SR) techniques to downscale coarse climate model outputs to higher resolutions. By treating climate data as images and employing stacked SRCNNs with auxiliary topographical data, DeepSD significantly improves the accuracy and scalability of climate change projections, outperforming existing statistical downscaling methods.

What Problem Does This Paper Address?

Climate change poses significant threats to various societal systems, from infrastructure to agriculture. However, current Earth System Models (ESMs) operate at spatial resolutions too coarse to capture localized impacts. While statistical downscaling (SD) techniques exist to bridge this gap, they often struggle with spatial dependencies and require extensive observational data, limiting their application in data-scarce regions. This paper addresses the challenge of generating high-resolution climate projections that are both accurate and scalable, particularly for regions with limited observational data, by adapting advanced image processing techniques.

Methods

The core innovation is the application of Single Image Super-Resolution (SR) techniques, specifically Convolutional Neural Networks (CNNs), to statistical downscaling. The proposed framework, DeepSD, is an augmented, stacked SRCNN architecture.

  • Stacked SRCNNs: Instead of directly upscaling from a low resolution (e.g., 1 degree) to a high resolution (e.g., 1/8 degree) in one step, DeepSD uses a series of SRCNNs, each performing a smaller resolution enhancement. This allows the model to learn spatial patterns at multiple scales, reducing complexity and improving representation.
  • Auxiliary Data Integration: Unlike traditional SR which only uses a low-resolution image, DeepSD incorporates static high-resolution topographical features (like elevation and land/water masks) as auxiliary input channels. This helps to constrain the downscaling process and improve accuracy, especially in regions with limited climate observations.
  • End-to-End Framework: The SRCNNs are trained independently for each scaling factor, and then stacked for inference. The process starts with the lowest resolution input and associated topographical data to predict the first resolution enhancement, and subsequent SRCNNs use the previous layer's output and corresponding topographical data.

Datasets

  • PRISM Dataset: Used for daily precipitation observations at a 4km spatial resolution. This dataset was upscaled to 1/8 degree (12.5km) for high-resolution evaluation and downscaled to 1 degree for low-resolution input.
  • GTOPO30 Elevation Dataset: A global digital elevation model with a 30 arcsec spatial resolution, used as static topographical input.

The paper does not introduce new benchmark datasets but rather uses existing observational data to train and evaluate its SD model, which can then be applied to ESM outputs.

Key Insights and Ideas

  • Climate Data as Images: The paper successfully demonstrates that treating climate data as images and applying computer vision techniques like SRCNNs can effectively capture spatial dependencies, a challenge for traditional SD methods.
  • Stacked SRCNNs for Scalability: The stacked approach allows for larger resolution enhancements (e.g., 8x or more) with greater accuracy than a single SRCNN, by breaking down the problem into manageable steps.
  • Topographical Data as a Powerful Proxy: Incorporating static topographical features significantly aids in downscaling, particularly in regions where direct climate observations are sparse, making the method more robust and applicable globally.
  • Potential for Improved Extreme Event Prediction: The authors suggest that DeepSD's ability to model extremes, beyond just matching quantiles, hints at improved prediction of extreme weather events.

Key Results

  • Performance Improvement: DeepSD outperformed other statistical downscaling methods (Lasso, ANN, SVM, SRCNN) and was closely behind BCSD in skill, while showing superior bias, correlation, and RMSE.
  • Scalability: DeepSD demonstrated significantly faster runtimes for downscaling compared to traditional methods like SVM, making it suitable for processing large ESM ensembles.
  • RMSE Reduction: DeepSD achieved a lower RMSE than BCSD for 79% of the CONUS region.
  • Generalization: The framework shows promise in generalizing beyond grid-by-grid predictions, suggesting it can leverage cross-grid information.
  • NEX Platform Integration: The paper discusses a framework for using DeepSD on the NASA Earth Exchange (NEX) platform to downscale multiple ESMs with various emission scenarios, enabling scalable generation of high-resolution climate products.

Limitations and Future Work

  • Spatial Generalization: The current experimentation is limited to regions where the model was trained. Future work should test predictability in regions with few observations.
  • Temporal Non-stationarity: The paper acknowledges that temporal non-stationarity (changes in climate patterns over time) is a challenge for statistical downscaling, and DeepSD's ability to handle this needs further evaluation.
  • Inclusion of More Variables: Incorporating additional climate variables (temperature, wind, humidity) and downscaling multiple variables simultaneously could improve the model's ability to capture complex climate patterns.
  • Uncertainty Quantification: A key limitation is DeepSD's failure to capture uncertainty in its downscaled projections. Future work should explore Bayesian deep learning concepts to address this.

Summary updated

BibTeX
@inproceedings{vandal2018generating,
  author = {Thomas Vandal and Evan Kodra and Sangram Ganguly and Andrew R. Michaelis and Ramakrishna R. Nemani and Auroop R. Ganguly},
  title = {{Generating High Resolution Climate Change Projections through Single Image Super-Resolution: An Abridged Version}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/759},
}
Figure