High-resolution ground NO2 estimation at hyperlocal level using deep learning with Sentinel-2 and Sentinel-5P data

Nitrogen dioxide (NO 2 ) is a harmful air pollutant that can cause various health issues, including respiratory disease and lung infection. Monitoring of NO 2 is primarily dependent on expensive ground-based sensor systems. This research explores the potential of integrating imagery from Sentinel-2 and Sentinel-5P to estimate high-resolution ground NO 2 concentration at city and neighborhood levels. This study presents a two-stream deep learning model for NO 2 estimation. The model is flexible regarding data input, allowing the use of Sentinel-2 and Sentinel-5P in combination or as single inputs from either satellite. The model performance is assessed over Chicago using Microsoft Eclipse ground sensor data aggregated in three temporal frequencies: daily, monthly, and quarterly. The experimental results demonstrate that fusing both satellite sources outperforms single-source models, achieving R 2 = 0.66, MSE = 5.92, and MAE = 1.75 at the quarterly scale, compared to R 2 = 0.59 for Sentinel-2 only and R 2 = 0.31 for Sentinel-5P only models. The estimated NO 2 is found to be most reliable at the quarterly level, followed by the monthly. Performance decreases at finer temporal scales (R 2 = 0.61 daily), likely due to the short-term fluctuation of NO 2 concentration. This study reinforces the application of deep learning and remote sensing for air quality monitoring, especially in the absence of expensive ground sensor-based monitoring systems. • Enhanced deep learning model for high-resolution NO 2 estimation. • Model evaluated at daily, monthly, and quarterly temporal scales. • Achieved best performance at the quarterly scale (R 2 = 0.66, MSE = 5.92, MAE = 1.75).