Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continuous latent representations for modeling precipitation with deep learning

Gokul Radhakrishnan, Rahul Sundar, Nishant Parashar, Antoine Blanchard, Daiwei Wang, Boyko Dodov

TL;DR

The paper tackles the challenge of modeling precipitation, which is sparse and heavily influenced by extremes, by introducing pseudo-precipitation ($PP$), a Gaussian-like latent field formed by blending total precipitation ($TP$) with the vertical integral of moisture divergence ($VIMD$) via an encoder–decoder and a quantile loss to enforce normality. This latent $PP$ is integrated into a downscaling framework that uses a spatio-temporal model (SimVP) and a diffusion-based decoder to recover high-resolution $TP$ at $0.25^ ext{}^\circ$ from low-resolution inputs, avoiding Gibbs-related artifacts associated with $TP$-only downscaling. The approach demonstrates that $PP$ preserves large-scale structure while enabling accurate reconstruction of fine-scale features and extremes, validated on ERA5 reanalysis data. Practically, this latent representation supports robust downscaling, debiasing, and forecasting, with potential for smoother assimilation of precipitation with other climate variables in impact studies.

Abstract

The sparse and spatio-temporally discontinuous nature of precipitation data presents significant challenges for simulation and statistical processing for bias correction and downscaling. These include incorrect representation of intermittency and extreme values (critical for hydrology applications), Gibbs phenomenon upon regridding, and lack of fine scales details. To address these challenges, a common approach is to transform the precipitation variable nonlinearly into one that is more malleable. In this work, we explore how deep learning can be used to generate a smooth, spatio-temporally continuous variable as a proxy for simulation of precipitation data. We develop a normally distributed field called pseudo-precipitation (PP) as an alternative for simulating precipitation. The practical applicability of this variable is investigated by applying it for downscaling precipitation from \(1\degree\) (\(\sim\) 100 km) to \(0.25\degree\) (\(\sim\) 25 km).

Continuous latent representations for modeling precipitation with deep learning

TL;DR

The paper tackles the challenge of modeling precipitation, which is sparse and heavily influenced by extremes, by introducing pseudo-precipitation (), a Gaussian-like latent field formed by blending total precipitation () with the vertical integral of moisture divergence () via an encoder–decoder and a quantile loss to enforce normality. This latent is integrated into a downscaling framework that uses a spatio-temporal model (SimVP) and a diffusion-based decoder to recover high-resolution at from low-resolution inputs, avoiding Gibbs-related artifacts associated with -only downscaling. The approach demonstrates that preserves large-scale structure while enabling accurate reconstruction of fine-scale features and extremes, validated on ERA5 reanalysis data. Practically, this latent representation supports robust downscaling, debiasing, and forecasting, with potential for smoother assimilation of precipitation with other climate variables in impact studies.

Abstract

The sparse and spatio-temporally discontinuous nature of precipitation data presents significant challenges for simulation and statistical processing for bias correction and downscaling. These include incorrect representation of intermittency and extreme values (critical for hydrology applications), Gibbs phenomenon upon regridding, and lack of fine scales details. To address these challenges, a common approach is to transform the precipitation variable nonlinearly into one that is more malleable. In this work, we explore how deep learning can be used to generate a smooth, spatio-temporally continuous variable as a proxy for simulation of precipitation data. We develop a normally distributed field called pseudo-precipitation (PP) as an alternative for simulating precipitation. The practical applicability of this variable is investigated by applying it for downscaling precipitation from ( 100 km) to ( 25 km).

Paper Structure

This paper contains 11 sections, 2 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Pseudo-precipitation (PP)
  4. Limitations of TP for downscaling
  5. Overview of the downscaling framework
  6. Results
  7. Experimental protocol
  8. Validation
  9. Discussion
  10. Appendix
  11. Training strategy

Figures (5)

  • Figure 1: Schematic of the mapping model for PP
  • Figure 2: Low and high resolution pairs of TP (left) and PP (right).
  • Figure 3: Visual assessment of downscaling for a snapshot. Downscaled PP is decoded to TP using the PP model decoder from figure \ref{['fig:schematics']}.
  • Figure 4: a) Power spectral density (PSD) and b) Q-Q plots.
  • Figure 5: Number of days of extreme precipitation (TP > 20 mm).