Enhancing Identifiable Variational Autoencoder in the Presence of Missing Values and Auxiliary Covariates for Spatio‐Temporal Ozone Modeling

Abstract

The formation of ground‐level ozone follows complex nonlinear photochemical processes that depend on multiple environmental factors and have strong spatio‐temporal structures. Environmental data used to study these dynamics usually originate from multiple sources, including in situ monitoring stations and satellite observations. While in situ data provide more accurate measurements, they often suffer from missing values and limited coverage, making it beneficial to incorporate additional satellite‐based covariates. To address these challenges for spatio‐temporal interpolation and forecasting aims, novel identifiable variational autoencoder (iVAE)‐based methods are introduced that explicitly integrate satellite‐derived predictors and handle missing values within a nonlinear blind source separation framework. The proposed extensions preserve the identifiability guarantees of the iVAE framework under the missing at random assumption. Performance is benchmarked against established statistical and deep learning approaches by using daily average ozone concentrations in Northern Italy: in interpolation, the proposed method achieves an appreciable reduction of the estimation errors and in forecasting it shows 10 times less training time with respect to the best competing method. The results establish nonlinear blind source separation as a promising approach for spatio‐temporal prediction of environmental data.