Abstract:
Global Climate Models (GCMs) simulate low-resolution climate projections on a global
scale. The native resolution of GCMs is generally too low for societal-level decisionmaking. Downscaling is often applied to GCM output to enhance the spatial resolution.
Statistical downscaling techniques, in particular, are well-established as a cost-effective
approach. They require significantly less computational time than physics-based dynamical downscaling. In recent years, deep learning has gained prominence in statistical
downscaling, demonstrating significantly lower error rates compared to traditional statistical methods. However, regression-based deep learning techniques, in particular, tend
to overfit the mean sample magnitude. As a result, extreme values are often underestimated. Furthermore, regression-based methods characteristically over-smooth the areas
surrounding sample extremes. The exact location of the sample extremes subsequently
becomes ambiguous. Problematically, extreme events have the largest societal impact,
e.g. substantial damage to infrastructure.
We propose Quantile-Regression-Ensemble (QRE), an innovative deep learning algorithm inspired by boosting methods. Its primary objective is to avoid trade-offs between
fitting to sample means and extreme values by training independent models on a partitioned dataset. Our QRE is robust to redundant models and not susceptible to explosive
ensemble weights, ensuring a reliable training process. Diffusion models can capture
high-frequency details of complex spatial patterns, providing a potential solution to over
smoothing. However, there is limited research on utilising their generative capacity in a
regression task, where sample magnitude may vary drastically. We introduce the SemDiff framework, which uses transfer learning on a surrogate dataset to train ScaleNet, a
network that scales diffusion samples to match the magnitude of the ground truth.
QRE achieves a lower Mean Squared Error (MSE) than various baseline models. In
particular, our algorithm has a lower error for high-intensity precipitation events over
New Zealand, highlighting the ability to represent extreme events reliably. SemDiff exhibits comparable capture of sample magnitude over baseline techniques while achieving
significantly higher SSIM and PSD. Therefore, Our proposed methods allow for reduced
uncertainty over extreme events, particularly when downscaling GCM projections.