Benchmarking the geographic generalization of deep learning models for precipitation downscaling

Area of Science:

• Climate Science
• Machine Learning
• Geospatial Analysis

Background:

• Earth System Models (ESM) are crucial for climate change projections but lack local-scale resolution.
• Deep learning super-resolution offers a solution for downscaling ESM outputs, but regional variations necessitate retraining.
• High-resolution observational data for retraining is unevenly distributed globally, creating accessibility inequities.

Purpose of the Study:

• To introduce RainShift, a dataset and benchmark for evaluating downscaling model generalization across geographic distribution shifts.
• To assess the performance of state-of-the-art downscaling models (GANs, diffusion models) under data gaps between the Global North and Global South.
• To investigate methods for improving the spatial generalization of climate model downscaling.

Main Methods:

• Developed the RainShift dataset and benchmark for evaluating geographic generalization.
• Evaluated Generative Adversarial Networks (GANs) and diffusion models on the RainShift benchmark.
• Assessed the impact of training domain expansion and domain adaptation techniques.

Main Results:

• Significant performance drops were observed in out-of-distribution regions, varying by model and location.
• Expanding training domains improved generalization but did not fully overcome shifts between geographically distinct regions.
• Domain adaptation techniques demonstrated potential for enhancing spatial generalization.

Conclusions:

• Current deep learning downscaling models exhibit limited generalization across diverse geographic regions.
• Addressing geographic distribution shifts is crucial for the global applicability of downscaling methods.
• Improving generalization can help reduce inequities in access to high-resolution climate information.