Detecting Changes in Power Plant NOx Emissions Using Deep Learning and TEMPO Satellite Imagery
Abstract
NASA's Tropospheric Emissions: Monitoring of Pollution (TEMPO) instrument provides hourly daytime observations of tropospheric nitrogen dioxide (NO2) across North America. These observations could support independent, facility-level monitoring of power plant nitrogen oxides (NOx) emissions, filling gaps when continuous emissions monitoring system (CEMS) measurements fail or facilities are not subject to hourly emissions reporting requirements. Previous studies have estimated hourly point-source emissions using inversion methods, which are highly sensitive to instrument retrieval uncertainty, changing atmospheric conditions, TEMPO's coarse spatial resolution, signals from nearby emission sources, and temporal mismatch with hourly CEMS records. We instead formulate facility monitoring as change detection over sequences of TEMPO observations. This allows the model to learn relative emissions behavior rather than estimate exact emissions from noisy data. We develop a convolutional gated recurrent unit (ConvGRU) model that processes five consecutive TEMPO observations together with meteorological inputs, features representing plume transport, and tabular features encoding time of day and seasonality to classify power plant emission trajectories as increasing, steady, or decreasing. On a held-out evaluation set of more than 22,000 samples covering 116 active U.S. power plants observed between August 2023 and June 2026, the ConvGRU model achieved a macro-F1 score of 0.56, outperforming a baseline multilayer perceptron (MLP) trained only on the same time-derived features, which scored 0.45. Our findings suggest that the ConvGRU model can identify signals of power plant emissions from TEMPO observations beyond the time-derived features shared with the baseline MLP. More broadly, they indicate that change detection may offer a more tractable framework for modeling power plant activity from TEMPO observations than estimating absolute emissions.
Related
- Spatial context importance in deep learning for global horizontal irradiance estimation from satellite imagery
- A Deep‐Learning Model Cascade for Automating Large‐Scale Shoreline Mapping From Satellite Imagery
- SATFRONT: Detecting Atmospheric Fronts with Satellite Data using Machine Learning
- Spatio-Temporal Analysis of Urban Expansion in Awka and its Environs, Nigeria Using Sentinel-2 Satellite Imagery and Google Earth Engine (2021 to 2026)
- Recurrent High-Resolution Satellite Observations Quantify Facility-Resolved Annual Large-Emitter Methane Emissions in the Permian Basin
BibTeX
@misc{walimbe2026detecting,
title = {{Detecting Changes in Power Plant NOx Emissions Using Deep Learning and TEMPO Satellite Imagery}},
author = {Pranav Walimbe and Ronald C Cohen and Hikari Murayama},
howpublished = {EarthArXiv},
year = {2026},
month = oct,
doi = {10.31223/x50j7k},
url = {https://eartharxiv.org/repository/view/15747/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Detecting Changes in Power Plant NOx Emissions Using Deep Learning and TEMPO Satellite Imagery
AU - Pranav Walimbe
AU - Ronald C Cohen
AU - Hikari Murayama
PY - 2026
DA - 2026/10/25
PB - EarthArXiv
DO - 10.31223/x50j7k
UR - https://eartharxiv.org/repository/view/15747/
AB - NASA's Tropospheric Emissions: Monitoring of Pollution (TEMPO) instrument provides hourly daytime observations of tropospheric nitrogen dioxide (NO2) across North America. These observations could support independent, facility-level monitoring of power plant nitrogen oxides (NOx) emissions, filling gaps when continuous emissions monitoring system (CEMS) measurements fail or facilities are not subject to hourly emissions reporting requirements. Previous studies have estimated hourly point-source emissions using inversion methods, which are highly sensitive to instrument retrieval uncertainty, changing atmospheric conditions, TEMPO's coarse spatial resolution, signals from nearby emission sources, and temporal mismatch with hourly CEMS records. We instead formulate facility monitoring as change detection over sequences of TEMPO observations. This allows the model to learn relative emissions behavior rather than estimate exact emissions from noisy data. We develop a convolutional gated recurrent unit (ConvGRU) model that processes five consecutive TEMPO observations together with meteorological inputs, features representing plume transport, and tabular features encoding time of day and seasonality to classify power plant emission trajectories as increasing, steady, or decreasing. On a held-out evaluation set of more than 22,000 samples covering 116 active U.S. power plants observed between August 2023 and June 2026, the ConvGRU model achieved a macro-F1 score of 0.56, outperforming a baseline multilayer perceptron (MLP) trained only on the same time-derived features, which scored 0.45. Our findings suggest that the ConvGRU model can identify signals of power plant emissions from TEMPO observations beyond the time-derived features shared with the baseline MLP. More broadly, they indicate that change detection may offer a more tractable framework for modeling power plant activity from TEMPO observations than estimating absolute emissions.
ER -