Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DiffObs: Generative Diffusion for Global Forecasting of Satellite Observations

Jason Stock, Jaideep Pathak, Yair Cohen, Mike Pritchard, Piyush Garg, Dale Durran, Morteza Mardani, Noah Brenowitz

TL;DR

DiffObs addresses the challenge of globally forecasting daily precipitation by training an autoregressive diffusion model on satellite-derived data to produce probabilistic day-ahead forecasts and stable multi-month rollouts. The method estimates $p(\mathbf{x}_{t} | \mathbf{x}_{t-1})$ without priors, using a conditioned diffusion process built on an EDM-inspired architecture. Results show DiffObs reproduces key tropical variability features, including the Madden–Julian Oscillation and convectively coupled Kelvin waves, as evidenced by Hovmöller analyses and Wheeler–Kiladis spectra, though artifacts such as dateline discontinuities and excess variance remain. The work demonstrates the potential of diffusion models trained on increasingly sparse and differentiated observations for subseasonal and climate prediction, while outlining future avenues to improve temporal distributions and cross-validated comparisons with reanalysis data.

Abstract

This work presents an autoregressive generative diffusion model (DiffObs) to predict the global evolution of daily precipitation, trained on a satellite observational product, and assessed with domain-specific diagnostics. The model is trained to probabilistically forecast day-ahead precipitation. Nonetheless, it is stable for multi-month rollouts, which reveal a qualitatively realistic superposition of convectively coupled wave modes in the tropics. Cross-spectral analysis confirms successful generation of low frequency variations associated with the Madden--Julian oscillation, which regulates most subseasonal to seasonal predictability in the observed atmosphere, and convectively coupled moist Kelvin waves with approximately correct dispersion relationships. Despite secondary issues and biases, the results affirm the potential for a next generation of global diffusion models trained on increasingly sparse, and increasingly direct and differentiated observations of the world, for practical applications in subseasonal and climate prediction.

DiffObs: Generative Diffusion for Global Forecasting of Satellite Observations

TL;DR

DiffObs addresses the challenge of globally forecasting daily precipitation by training an autoregressive diffusion model on satellite-derived data to produce probabilistic day-ahead forecasts and stable multi-month rollouts. The method estimates without priors, using a conditioned diffusion process built on an EDM-inspired architecture. Results show DiffObs reproduces key tropical variability features, including the Madden–Julian Oscillation and convectively coupled Kelvin waves, as evidenced by Hovmöller analyses and Wheeler–Kiladis spectra, though artifacts such as dateline discontinuities and excess variance remain. The work demonstrates the potential of diffusion models trained on increasingly sparse and differentiated observations for subseasonal and climate prediction, while outlining future avenues to improve temporal distributions and cross-validated comparisons with reanalysis data.

Abstract

This work presents an autoregressive generative diffusion model (DiffObs) to predict the global evolution of daily precipitation, trained on a satellite observational product, and assessed with domain-specific diagnostics. The model is trained to probabilistically forecast day-ahead precipitation. Nonetheless, it is stable for multi-month rollouts, which reveal a qualitatively realistic superposition of convectively coupled wave modes in the tropics. Cross-spectral analysis confirms successful generation of low frequency variations associated with the Madden--Julian oscillation, which regulates most subseasonal to seasonal predictability in the observed atmosphere, and convectively coupled moist Kelvin waves with approximately correct dispersion relationships. Despite secondary issues and biases, the results affirm the potential for a next generation of global diffusion models trained on increasingly sparse, and increasingly direct and differentiated observations of the world, for practical applications in subseasonal and climate prediction.
Paper Structure (13 sections, 8 equations, 7 figures)

This paper contains 13 sections, 8 equations, 7 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Training Details
  4. Experiments
  5. Dataset Details
  6. Main Findings
  7. Conclusion
  8. Diffusion Details
  9. Wheeler--Kiladis Space-Time Spectra
  10. Short-Term Predictability
  11. Additional Experiments
  12. Modifications
  13. Preliminary Results

Figures (7)

  • Figure 1: Example 3-day rollout from Oct 27, 2020 as initial condition.
  • Figure 2: Hovmöller diagrams of observations (left) and DiffObs output (right, generated autoregressively) between $5^\circ$N and $5^\circ$S for case studies initially conditioned on (a) July 1, 2020 and (b) Oct 1, 2019. Individual colors correspond to the wave propagation directions (W$\leftrightarrow$E), Indian Summer Monsoon (ISM), Madden‐-Julian oscillation (MJO), and East Pacific Wavetrain (EPW).
  • Figure 3: Symmetric / Background Wheeler--Kiladis space-time spectra between $15^\circ$N and $15^\circ$S. The individually highlighted regions correspond to where the Madden‐-Julian oscillation (MJO), Kelvin and westward inertio-gravity (WIG) waves are expected to be found.
  • Figure 4: Reverse diffusion of a cropped sample with the input condition, individual sampling steps ($t_0\rightarrow t_{64}$, inversely labeled), and the next time step estimate and target output.
  • Figure 5: Additional Wheeler--Kiladis components and power spectra of observations (left column) and model output (right column) that support \ref{['fig:results-wk']}.
  • ...and 2 more figures