Spatial emissions modeling from mobile sources using hidden Markov models and autoencoders

Abstract

Air pollution poses a worldwide threat to climate, health, and ecosystems, particularly in urban regions where mobile sources contribute significantly. The magnitude and spread of these emissions are influenced by multiple spatial factors. This study aims to classify pollutants from mobile sources by independently analyzing the spatial characteristics of their trajectory. An effective approach to evaluating environmental factors and their impact on emissions across various locations is essential for air pollution management. This research focuses on categorizing emissions produced by single taxi units using sequential models such as Hidden Markov Models (HMM), along with advanced hybrid techniques that combine HMM with Autoencoders (AE) and Variational Autoencoders (VAE). While HMMs capture temporal dependencies in sequential data, they assume low-dimensional, noise-free observations—an assumption often violated by real-world emissions data. To address this, AE and VAE are used to transform high-dimensional inputs into structured latent representations that align with HMM assumptions. Given the presence of various sources of error, this study also investigates the uncertainty associated with each method. Results show that the HMM-VAE hybrid outperforms other methods in classification, achieving test data accuracies of 0.80 for fuel consumption (FC), 0.79 for nitrogen oxides (Nox), 0.78 for hydrocarbons (HC), and 0.76 for carbon monoxide (CO). Uncertainty assessment reveals this hybrid method has lower uncertainty and higher confidence. Feature removal analysis showed that removing transportation hubs and taxi counts significantly decreased HMM accuracy. This study demonstrates that hybrid sequential-autoencoder models accurately map mobile source emissions, aiding air pollution risk assessment in large-scale monitoring.