Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Likelihood-Based Generative Approach for Spatially Consistent Precipitation Downscaling

Jose González-Abad

TL;DR

The paper tackles spatial inconsistency in precipitation downscaling when sampling from likelihood-based distributions. It introduces a likelihood-based generative approach that couples a conditional GAN with a Bernoulli-gamma distribution to produce spatially coherent precipitation fields while delivering an explicit probabilistic output. Key contributions include the cGAN (NLL) model, which uses the Bernoulli-gamma $NLL$ as the content loss, and a comparative analysis against U-Net baselines and a cGAN (MSE). Results show improved representation of extreme events and spatial coherence, with the cGAN (NLL) balancing realism and distribution fidelity. This approach provides a probabilistic downscaling framework with practical relevance for climate risk assessment and future GCM evaluations.

Abstract

Deep learning has emerged as a promising tool for precipitation downscaling. However, current models rely on likelihood-based loss functions to properly model the precipitation distribution, leading to spatially inconsistent projections when sampling. This work explores a novel approach by fusing the strengths of likelihood-based and adversarial losses used in generative models. As a result, we propose a likelihood-based generative approach for precipitation downscaling, leveraging the benefits of both methods.

A Likelihood-Based Generative Approach for Spatially Consistent Precipitation Downscaling

TL;DR

The paper tackles spatial inconsistency in precipitation downscaling when sampling from likelihood-based distributions. It introduces a likelihood-based generative approach that couples a conditional GAN with a Bernoulli-gamma distribution to produce spatially coherent precipitation fields while delivering an explicit probabilistic output. Key contributions include the cGAN (NLL) model, which uses the Bernoulli-gamma as the content loss, and a comparative analysis against U-Net baselines and a cGAN (MSE). Results show improved representation of extreme events and spatial coherence, with the cGAN (NLL) balancing realism and distribution fidelity. This approach provides a probabilistic downscaling framework with practical relevance for climate risk assessment and future GCM evaluations.

Abstract

Deep learning has emerged as a promising tool for precipitation downscaling. However, current models rely on likelihood-based loss functions to properly model the precipitation distribution, leading to spatially inconsistent projections when sampling. This work explores a novel approach by fusing the strengths of likelihood-based and adversarial losses used in generative models. As a result, we propose a likelihood-based generative approach for precipitation downscaling, leveraging the benefits of both methods.
Paper Structure (11 sections, 3 figures)

This paper contains 11 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Background
  3. Deep Learning for Precipitation Downscaling
  4. Extreme Precipitation
  5. Generative Precipitation Downscaling
  6. Experimental Framework
  7. Region of Study and Data
  8. Standard Deep Learning Models
  9. Likelihood-Based Generative Approach
  10. Results
  11. Conclusions

Figures (3)

  • Figure 1: Violin plot showing the results for four different metrics computed on the test set: the relative bias of the mean and the SDII, the RMSE, and the ratio of standard deviations. Each metric displays the results corresponding to the different DL models intercompared: U-Net (MSE), U-Net (NLL), cGAN (MSE), and cGAN (NLL).
  • Figure 2: Histogram of the precipitation distribution for the test period, aggregated across all grid-points in the predictand. The black line represents the target observational dataset, while the different colors correspond to the various DL models being compared. A zoomed-in view for values in the 0-50 interval is provided in the top-right corner of the histogram.
  • Figure 3: Comparison of predictions generated by the DL models intercompared for a day in the test period.