Spatiotemporal Gap-Filling of NASA Deep Blue Satellite Aerosol Optical Depth Over the Contiguous United States (CONUS) Using the UNet 3+ Architecture

Abstract

Due to sensor and algorithmic constraints, satellite aerosol optical depth (AOD) retrievals are spatially incomplete and have gaps caused by clouds and bright surfaces. These gaps represent a barrier in characterizing daily aerosol loadings, which is important for air quality applications. In particular, recent studies in aerosol studies have shown satellite AOD to be a useful predictor of particulate matter, but are often limited to monthly or longer temporal resolution because of missing AOD retrievals. In this study, we propose using a UNet 3+ to fill gaps in satellite AOD retrievals. We tested the hypothesis that UNet 3+ trained on deep blue (DB) AOD and supplemental data sets (e.g., Modern-Era Retrospective analysis for Research and Applications, Version 2 reanalysis AOD, meteorological and land-use variables from North American Mesoscale Forecast System, and Hazard Mapping System smoke polygons) will improve the availability of AOD data accurately. We created spatiotemporal data sets of daily, gap-filled DB AOD from 2012 to 2023 over the CONtinental United States (CONUS) at a 12 × 12 km² resolution. We were able to train the model and perform the gap-filling in ∼10 hr, resulting in an increase of AOD data availability by 281%. We demonstrated that our approach is feasible over CONUS through quantitative and qualitative evaluations against AERONET and DB AOD. In statistical evaluations, our gap-filled AOD data set attained an RMSE ∼ 0.09 and a r ∼ 0.87 against collocated AERONET retrievals, compared to an RMSE ∼ 0.11 and a r ∼ 0.86 that the original DB AOD retrievals scored against AERONET. We plan to use this data set for future air quality and health investigations.