AODDiff: Probabilistic Reconstruction of Aerosol Optical Depth via Diffusion-based Bayesian Inference

TL;DR

AODDiff tackles the problem of reconstructing high-resolution AOD fields from incomplete satellite observations by reframing it as a conditional diffusion-based Bayesian inference task. It learns a spatiotemporal AOD prior from incomplete data using corruption-aware training and employs a decoupled annealing posterior sampling (DAPS) strategy to flexibly fuse heterogeneous observations without retraining. The approach yields superior spectral fidelity and provides explicit uncertainty quantification through multiple posterior samples, validated on MERRA-2 reanalysis data for downscaling and inpainting tasks. This framework enables scalable, uncertainty-aware reconstruction of gap-filled AOD fields, with potential to improve aerosol monitoring and related risk assessments.

Abstract

High-quality reconstruction of Aerosol Optical Depth (AOD) fields is critical for Atmosphere monitoring, yet current models remain constrained by the scarcity of complete training data and a lack of uncertainty quantification.To address these limitations, we propose AODDiff, a probabilistic reconstruction framework based on diffusion-based Bayesian inference. By leveraging the learned spatiotemporal probability distribution of the AOD field as a generative prior, this framework can be flexibly adapted to various reconstruction tasks without requiring task-specific retraining. We first introduce a corruption-aware training strategy to learns a spatiotemporal AOD prior solely from naturally incomplete data. Subsequently, we employ a decoupled annealing posterior sampling strategy that enables the more effective and integration of heterogeneous observations as constraints to guide the generation process. We validate the proposed framework through extensive experiments on Reanalysis data. Results across downscaling and inpainting tasks confirm the efficacy and robustness of AODDiff, specifically demonstrating its advantage in maintaining high spatial spectral fidelity. Furthermore, as a generative model, AODDiff inherently enables uncertainty quantification via multiple sampling, offering critical confidence metrics for downstream applications.

Figures (9)

• Figure 1: Spatiotemporal Scope and Train/Validation Split of the AOD Dataset
• Figure 2: Workflow of the observation operators. The reference field $\mathbf{X}_{\text{GT}}$ is processed by the masking operator $\mathcal{A}_{Mask}$ to simulate cloud-occluded gaps, and the downsampling operator $\mathcal{A}_{DS}$ to generate low-resolution representations.
• Figure 3: Overall framework of the AODDiff: (a) The corruption-aware training process, where the score network learns the spatiotemporal prior from incomplete observations. (b) The inference process, which employs decoupled annealing posterior sampling strategy to reconstruct the AOD field.
• Figure 4: Comparison of (a) Spatial Mean and (b) Standard Deviation of Ground Truth and Generated Sample Sets.
• Figure 5: Comparison of (a) Rotational Average Power Spectral Density (RAPSD) and (b) Spatially Averaged Temporal Autocorrelation (ACF) across Ground Truth and Generated Sample Sets.