Abstract
Precipitation downscaling is critical for predicting local climate impacts, yet models often struggle to generalize across different geographic regions. We present RainShift, a benchmark that evaluates the performance of state-of-the-art deep learning models—including ResNet, WGANs, and Diffusion models—across diverse climates. Our findings reveal a significant "geographic gap": models trained on Global North data experience up to a 37% performance drop when applied to the Global South. This benchmark highlights that geographic variability, rather than model architecture alone, is the primary bottleneck for global climate equity in AI.
Introduction
Earth System Models (ESMs) provide essential climate predictions but operate at coarse resolutions (approx. 100 km), which are insufficient for local disaster planning. Deep learning-based "statistical downscaling" offers a high-resolution alternative. However, a major limitation in current research is the lack of cross-regional validation. High-resolution ground data is concentrated in the Global North, leaving the Global South underrepresented. RainShift addresses this by testing model generalization across 12 training and 6 evaluation regions worldwide.
Methodology
The RainShift benchmark evaluates three primary architectures against a bilinear interpolation baseline:
- ResNet: A deterministic model providing a single point estimate.
- Wasserstein GAN (WGAN): A probabilistic model that generates realistic precipitation textures.
- Diffusion Models: Generative models that iteratively refine noise into high-fidelity precipitation samples.
A critical component of the methodology is Quantile Mapping (QM), used to align input data distributions between training and evaluation regions, mitigating distribution shifts without retraining.
Datasets and Regional Distribution
To ensure global consistency, RainShift utilizes:
- Low-Resolution Input: ERA5 atmospheric reanalysis data (~25 km resolution).
- High-Resolution Target: IMERG satellite-based precipitation data (~10 km resolution).
The benchmark defines 12 training regions primarily in the Global North and 6 evaluation regions in the Global South (e.g., Amazon Basin, Horn of Africa). This setup tests "zero-shot" generalization capabilities.
Key Insights
Geography vs. Architecture
The specific climate of a region has a more profound impact on model performance than the choice of architecture. Highly variable climates, such as the Amazon, are inherently more difficult to predict for all models.
The Data Inequity Trap
Models trained exclusively on Global North data show substantial performance declines (up to 37%) in the Global South. This suggests that simply increasing data volume in the North does not solve the prediction problem for the South.
Model Robustness and Alignment
Diffusion models demonstrate greater robustness with limited training data, while GANs scale better with larger datasets. Furthermore, statistical alignment via Quantile Mapping proved essential; it allowed models that were previously unstable in new regions to produce accurate results.
Results and Future Work
All deep learning models achieved a 30% to 50% improvement in Continuous Ranked Probability Score (CRPS) over standard bilinear interpolation. Generative models (GANs and Diffusion) consistently outperformed deterministic approaches by capturing the stochastic nature of extreme rainfall.
Future research will focus on moving beyond zero-shot generalization toward few-shot adaptation, incorporating dynamic features like soil moisture, and addressing the inherent biases in satellite-based "ground truth" data.
Summary updated