Bayesian spatio–temporal disaggregation modeling using a diffusion-SPDE approach: a case study of Aerosol Optical Depth in India
Fernando Rodriguez Avellaneda · 2025
Read Bayesian spatio–temporal disaggregation modeling using a diffusion-SPDE approach: a case study of Aerosol Optical Depth in India by Fernando Rodriguez Avellaneda (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-06276.txt.
Introduction
Fernando Rodriguez Avellaneda and Paula Moraga sharpen estimates of aerosol optical depth, a measure of particles in the air tied to climate and health. Satellites report this quantity only at coarse scales in space and time. Finer maps would let officials target air quality policy and health interventions far more precisely.
Their Bayesian model treats the true field as a smooth process in space and time seen only through averaged readings. Covariances come from a diffusion-based partial differential equation, and fast approximate inference keeps the computation practical. In an India case study, the method tripled spatial detail and cut the time step from three hours to one.
Catalog record
| CREATOR | Fernando Rodriguez Avellaneda |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-06276
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Bayesian spatio–temporal disaggregation modeling using a diffusion-SPDE approach: a case study of Aerosol Optical Depth in India Fernando Rodriguez Avellaneda Abstract Accurate estimation of Aerosol Optical Depth (AOD) is crucial for understanding climate change and its impacts on public health, as aerosols are a measure of air quality conditions. AOD is usually retrieved from satellite imagery at coarse spatial and temporal resolutions. However, producing high-resolution AOD estimates in both space and time can better support evidence-based policies and interventions. We propose a spatio-temporal disaggregation model that assumes a latent spatio–temporal continuous Gaussian process observed through aggregated measurements. The model links discrete observations to the continuous domain and accommodates covariates to improve explanatory power and interpretability. The approach employs Gaussian processes with separable or non-separable covariance structures derived from a diffusion-based spatio-temporal stochastic partial differential equation (SPDE). Bayesian inference is conducted using the INLA-SPDE framework for computational efficiency. Simulation studies and an application to nowcasting AOD at 550 nm in India demonstrate the model’s effectiveness, improving spatial resolution from 0.75 ∘ 0.75^{\circ} to 0.25 ∘ 0.25^{\circ} and temporal resolution from 3 hours to 1 hour. Keywords: Air pollution; Disaggregation; Non–separable covariance; Spatio–temporal Modeling; Difussion SPDE; INLA 1 Introduction Aerosols, commonly quantified through Aerosol Optical Depth (AOD), are a fundamental component of atmospheric composition with wide-ranging implications for both climate and human health [ 19 ] . In particular, AOD at 550 nanometers (nm) is a dimensionless quantity that measures the degree to which aerosols (e.g., dust, smoke, haze) scatter or absorb light in a vertical column of the atmosphere, using the middle of the visible spectrum (550 nm) as the reference wavelength. AOD is commonly combined with chemical transport models and ground observations to estimate and forecast ground–level particulate matter, thereby supporting air quality monitoring [ 31 ] . Elevated AOD measurements are frequently associated with high risks of respiratory and cardiovascular conditions that could include asthma, chronic obstructive pulmonary disease (COPD), and ischemic heart disease [ 28 ] . Beyond health effects, aerosols also modify the Earth’s radiative balance, cloud microphysics, and precipitation patterns, thereby linking AOD to broader climate processes with negative impacts on ecosystems and agricultural productivity [ 25 ] . Understanding the spatio–temporal distribution of AOD is critical for informing policies aimed at reducing its adverse effects on human health, ecosystems, and food security. A significant obstacle, however, lies in the limited availability of high–-resolution AOD data; ground monitoring stations remain sparse in many parts of the world, while reanalysis products such as CAMS [ 8 ] or satellite retrievals from instruments such as MODIS [ 20 ] and MISR [ 9 ] provide global coverage only at coarse spatial (tens of kilometers) and temporal (3–-6 hours) resolutions, which are often insufficient for public health and environmental applications where fine-scale variability is crucial [ 5 , 24 ] . This mismatch between the spatial resolution of available data and the finer scales required is often referred as the change of support problem [ 18 , 21 ] . In this study, we develop a spatio-temporal modeling framework that enable the disaggregation of coarsely aggregated AOD data to produce high–resolution estimates in both space and time. We evaluate the performance of our model by downscaling AOD at 550 nm in India, a region charaterized by high aerosol loading, diverse topography, and significant public health relevance. Specifically, we use the Atmospheric Composition Reanalysis 4 (EAC4) dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF) [ 8 ] . The dataset provides AOD values on a spatial grid of 0.75^{\circ}\times 0.75^{\circ} in latitude and longitude with a temporal resolution of 3 hours. In this study, we perform spatio-temporal disaggregation to achieve a finer spatial resolution of 0.25^{\circ} and 1 hour in time, thereby providing a more detailed representation of aerosol variability that can support improved analyses of its impacts on air quality and health. The problem of disaggregation has been widely studied in the spatial framework, particularly in the context of air pollution [ 18 ] . Existing approaches can be broadly classified into geostatistical methods and machine learning models. Machine learning methods are often computationally efficient and can handle large datasets, but they usually treat covariates in a black-box manner and operate at aggregated levels, which limits interpretability and their ability to capture spatial dependence explicitly [ 14 ] . In contrast, geostatistical approaches leverage spatial autocorrelation to improve predictions and allow the explicit inclusion of covariates through Gaussian process models [ 26 , 30 ] . However, these models often become computationally prohibitive, as likelihood-based inference and prediction typically require costly approximations due to the lack of closed-form solutions [ 1 , 7 ] . Bayesian melding and downscaler models have been presented by several authors to combine spatially misaligned data [ 32 ] . For example, a Bayesian spatio-temporal downscaler was proposed by [ 2 ] , linking coarse numerical model outputs with point observations through spatially varying coefficients driven by Gaussian processes. [ 16 ] introduced a Bayesian melding that assumes an underlying continuous latent Gaussian Random Field (GRF) observed at an aggregated level and uses INLA and SPDE for fast Bayesian inference [ 13 ] , applied to map air pollution in California, USA. Building on this idea, [ 21 ] extended the approach to a multivariate setting, applying it in Portugal and Italy to disaggregate PM2.5, PM10, and Ozone concentrations at finer spatial resolutions. More recently, scalable developments such as DALIA [ 4 ] have addressed the problem of spatial disaggregation by utilizing efficient Bayesian inference for spatio-temporal multivariate Gaussian processes, thereby overcoming the computational bottlenecks of high-dimensional models through sparse precision formulations and parallelization strategies. While this approach used multivariate spatio-temporal data, its disaggregation is restricted to the spatial domain. In parallel, machine learning approaches have introduced attention-based models to disaggregate spatio–temporal data, typically from coarse administrative units (e.g., census tracts) to finer ones (e.g., city blocks). A notable example is the Structurally-Aware Recurrent Network (SARN), which integrates spatial attention layers with recurrent neural networks to jointly capture spatial interactions and temporal dependencies [ 6 ] . Previous models have mainly focused on spatial disaggregation, maintaining computational scalability, and to the best of our knowledge none have explicitly addressed the joint spatio–temporal disaggregation problem. In this paper, we propose a spatio–temporal disaggregation model for normally distributed data, applied here to aerosol optical depth (AOD) at 550 nm but broadly applicable to other spatio–-temporal contexts. Our model assumes the existence of a spatio–temporal latent Gaussian process with either a separable or non–separable covariance structure in space and time, providing the flexibility to capture realistic dependencies in space and time. Computational efficiency is achieved through the SPDE–GMRF framework, with inference performed using the INLA and INLAspacetime packages [ 22 , 11 ] . This approach directly addresses the challenge of linking data observed at aggregated levels to an underlying continuous latent process, enabling predictions at finer spatial and temporal resolutions. The remainder of the paper is organized as follows. Section 2 presents the methodological framework for the spatio-temporal disaggregation model, and the inference strategy adopted in this work. Section 3 reports the results of a simulation study evaluating the performance of our spatio–temporal disaggregation