Estimating and Forecasting Surface NO₂ from TEMPO via Machine Learning

Abstract

Nitrogen dioxide (NO₂) is a pollutant that is harmful to respiratory health, making it important to track and forecast. Traditional methods for this are limited, as there are only a few hundred EPA monitors across the contiguous United States, and forecasts typically require supercomputer-scale chemical transport simulations, which have biases relative to EPA monitor readings. To address these limitations, we present a three-stage machine-learning framework: the first stage converts TEMPO's L3 NO₂ vertical column density (VCD) observations into 0.1° surface-concentration (ppb) estimations, the second stage forecasts NO₂ surface concentration directly at each time t ∈ {1, 3, 6, 12, 24, 48} hours, and the third stage converts the forecast at every grid cell into a point-level estimate. This approach does not require ground sensors or a chemical simulation at runtime, allowing for computationally cheap, widespread NO₂ forecasting capabilities. First, a LightGBM model maps each TEMPO observation to surface concentrations using 20 features including meteorology features, temporal features, and static land-use proxy features (wind vector components, temperature, solar hour, nighttime lights, NDVI, etc.). The first stage attains an R² of 0.80 at monitored sites on six held-out months (October 2024–March 2025) and 0.51 in areas whose monitors were withheld from training. The second stage is a U-Net that takes the previous eight hours of surface maps and forecasts the full surface NO₂ concentration field directly at each forecast time. The third stage is a second LightGBM that predicts point-level concentrations from the U-Net's cell averages using local land-use proxies, meteorology, and time features. When tested at held-out times for stations seen during training, the model outperforms GEOS-CF at every forecast time, with an RMSE of 5.1 vs. 9.6 ppb at one hour and 5.8 vs. 10.2 ppb at 48 hours. At stations withheld from training, our model’s performance degrades but still beats GEOS-CF, achieving 7.8 vs. 9.3 ppb at one hour and 8.2 vs. 9.9 ppb at 48 hours. This shows that the model outperforms GEOS-CF whether or not EPA monitors exist at a location, at a fraction of the computational cost at inference.