Machine Learning for Air Quality Prediction Using High-Resolution TEMPO Remote Sensing Data

NASA's Tropospheric Emissions: Monitoring of Pollution (TEMPO) instrument provides advanced measurements of key pollutants—ozone, nitrogen dioxide, formaldehyde, sulfur dioxide, and aerosols—using an ultraviolet and visible-light grating spectrometer. From
geostationary orbit, TEMPO delivers high temporal and spatial resolution, with an hourly pixel resolution of 2 km (North/South) and 4.7 km (East/West) over North America. This fine-scale data enables near real-time tracking of pollution patterns on an urban scale,
offering insights into hourly air quality dynamics. We aim to utilize machine learning to predict tropospheric nitrogen dioxide (NO2) during the validation phase of the TEMPO project. These data should be considered as provisional products per the Provisional Product Maturity level defined in the TEMPO validation plan. These data are at provisional maturity, which means that product performance has been demonstrated through a large, but still (seasonally or otherwise) limited number of independent measurements. We initialize our training with Random Forests to predict in the absence of TEMPO measurements, validating with validated ground-based Pandora NO2 measurements, ensuring consistency with real-world observations. To further enhance predictions, we propose a hybrid approach integrating a Random Forest with a Convolutional Neural Network (CNN).