Nighttime PM_2.5 Concentration Estimation Based on NPP-VIIRS and Interpretable Machine Learning: The Case of Beijing–Tianjin–Hebei Region

Air pollution and public health issues caused by fine particulate matter (PM2.5) are becoming increasingly severe. Although well-established satellite methods exist for retrieving daytime PM2.5 concentrations, these methods are limited by weak nighttime light radiation. To resolve these challenges, this study proposed a nighttime PM2.5 concentration estimation method based on explainable machine learning and low-light data. Owing to the complexity of nighttime light sources, primarily composed of artificial lighting and moonlight, both types of light were considered by simulating lunar irradiance and artificial light radiance. This study utilized nighttime lighting, meteorological, various geospatial auxiliary, simulated nighttime light radiation, and ground-based PM2.5 monitoring data to construct a dataset with an effective sample size of 24,311. A deep neural network model was trained to estimate nighttime PM2.5 concentrations. Experimental results show that, after adding the simulated nighttime light radiation, the ten-fold cross-validation R2 of the model improved from 0.6 to 0.73. Additionally, 74% of site-based ten-fold cross-validation R2 values exceeded 0.7, indicating the model's robust spatial adaptability. The model was then used to estimate nighttime PM2.5 concentrations in the study area for 2021. The Shapley additive explanation (SHAP) model was applied to analyze the effect curves of different predictors on nighttime PM2.5 to examine the contributions of various factors. This study can serve as a reference for similar research in the future, and the proposed retrieval method offers a broad coverage of nighttime PM2.5 data, providing a useful supplement to ground station measurements.