Enhancing Prediction of Satellite-Based NO2 Concentration in ASEAN Using Random Forest Models with Spatial Coordinates and Temporal Lag Features

Abstract

Nitrogen dioxide (NO2) is a critical indicator of anthropogenic emissions, making its monitoring essential for environmental management in the rapidly urbanizing ASEAN region. While satellite imagery provides the necessary high spatial coverage to overcome the limitations of sparse ground-based stations, traditional machine learning models applied to these datasets often overlook the inherent spatial heterogeneity and temporal persistence of air pollutants. This study aims to quantify the marginal contributions of geographic coordinates and temporal lag features to the prediction accuracy of satellite-derived NO2 concentrations using Random Forest (RF) models. By evaluating four progressive RF configurations across 1,500 locations from July 2018 to December 2024, the study identifies the optimal feature combination for regional air quality modeling. The results demonstrated that the RF model incorporating both geographic coordinates and a 12-month lag variable achieved the best performance, yielding an  of 0.832 and an RMSE of 5.42 . Feature importance analysis revealed that the 12-month lag of NO2, nighttime lights, and location were the most influential predictors, highlighting the strong annual seasonality and the impact of economic activities on pollution levels. These findings provide a robust, data-driven framework for regional air quality monitoring and policy formulation in developing tropical regions.