Bridging Monitoring Gaps: Multi-Task Learning Enhances High-Resolution Surface-Level O 3 , NO 2 , and NO Estimation

Download PDF
Open Webpage

Zhongying Wang, James Crooks, Elizabeth Regan, Morteza Karimzadeh

Published: 09 Dec 2025, Last Modified: 07 May 2026CrossrefEveryoneRevisionsCC BY-SA 4.0
Abstract: Ground-level ozone (O 3 ), nitrogen dioxide (NO 2 ), and nitric oxide (NO) are key components of photochemical smog and well-established threats to public health and ecosystems. Because these gases interconvert within hours through photochemical reactions (e.g., NO + O 3 → NO 2 + O 2 ), modeling them together lets information from each pollutant inform and improve the prediction of the others. However, regulatory monitoring networks remain sparse and uneven across space and time: the Contiguous United States (CONUS) hosts approximately 1000 O 3 sites but only 400 NO 2 and fewer NO stations, posing a challenge to the development of data-hungry deep learning models. We present a multi-task learning (MTL) framework that simultaneously estimates daily 1 km-resolution O 3, NO 2, and NO concentrations for CONUS. The architecture couples a shared spatiotemporal encoder—composed of a convolutional neural network (CNN) followed by a bi-directional long short-term memory (Bi-LSTM) network—with learned location and temporal embeddings; pollutant-specific decoders then produce individual pollution concentrations. Training leverages a comprehensive set of predictors: meteorology from Daymet (temperature, precipitation, vapor pressure, short‑wave radiation) and GridMET (wind speed and direction); chemical precursors from the National Emissions Inventory (VOC and NO x ); land‑surface information from MODIS NDVI and a digital elevation model (DEM); reanalysis data from CAMS EAC4 and MERRA‑2; and satellite retrievals from Aura/OMI (total‑column O 3, NO 2, NO). MTL yields clear benefits over single-task baselines in all three pollutants, demonstrating that knowledge shared across chemically linked pollutants enhances generalization even for the well-monitored pollutant. The resulting 1km, daily O 3 -NO 2 -NO dataset from 2005 to 2021 will be made publicly available, providing an unprecedented resource for longitudinal epidemiological studies, exposure disparity assessments, and air quality management.