Hyper-local black carbon prediction by integrating land use variables with explainable machine learning model

Abstract

This research introduces an innovative framework that effectively integrates the Land Use Regression (LUR) model with an explainable machine learning methodology, thereby enhancing the causal understanding of how land use variables impact Black Carbon (BC) concentrations in complex urban landscapes. Through the development and application of a LUR-based Bayesian Network (BN) model to detailed, mobile-based BC measurements, we have successfully decoded the causal relationships and conditional probabilities linking diverse urban land use variables to BC levels. The BN model demonstrates notable prediction accuracy on the test set, with an R2 of 0.59 and an NRMSE of 0.65. Among the varied land use variables analyzed, proximity to the port is identified as the dominant factor contributing to BC hotspots. Moreover, the synergistic effect of port proximity and truck routes emerges as a key driver of spatially clustered BC hotspots, a complex phenomenon not adequately captured by conventional machine learning and deep learning models. The insights revealed through this study are vital not just for enhancing air pollution prediction and exposure analysis but also for informing urban planning and policymaking, offering valuable, actionable insights.