Integrating Machine Learning with Causal Inference to Improve Prediction of Ammonium Wet Deposition in the Pearl River Delta

Atmospheric nitrogen deposition is a vital component of the global nitrogen cycle, with significant implications for ecosystem health, pollution mitigation, and sustainable development. In the Pearl River Delta (PRD) region of China, high levels of ammonium ( NH x ) wet deposition, driven by abundant precipitation and intensive anthropogenic activities, pose significant challenges to ecological balance and environmental sustainability. However, accurately estimating NH x wet deposition flux is hindered by the complexity of nitrogen deposition processes and spatial heterogeneity in observational data. This study integrates machine learning and causal inference techniques to identify the spatial distribution patterns of NH x wet deposition and key drivers of its spatial heterogeneity. Based on these findings, four machine learning models were developed to estimate NH x wet deposition flux in the PRD region for the period 2012–2017. The results indicated that the integrated models significantly outperformed standard machine learning models (MSE = 0.486, R = 0.564), the FGCNN deep learning model (MSE = 0.454, R = 0.592), and the WRF-EMEP numerical model (MSE = 0.975, R = 0.334), achieving the highest average accuracy (MSE = 0.379, R = 0.610). This study emphasizes the importance of incorporating causal factors and spatial heterogeneity into estimation frameworks to improve the accuracy and stability of NH x wet deposition flux estimates. The findings provide actionable insights for targeted mitigation strategies, contributing to sustainable ecosystem management and pollution reduction in rapidly urbanizing regions.