A Coupled Deep Learning Model for Estimating Surface NO₂ Levels From Remote Sensing Data: 15-Year Study Over the Contiguous United States

This study proposes a novel two-step deep learning (DL) model for estimating surface NO2 concentrations using satellite data over the contiguous United States (CONUS) from 2005 to 2019. The first phase of the model uses partial convolutional neural network (PCNN), an advanced DL model that accurately imputes gaps between surface NO2 stations and creates 5,478 daily-mean NO2 grids (PCNN-NO2) of the 2005-2019 period over the study area. We then feed the PCNN-NO2, along with other predictor variables, into a deep neural network (DNN) to estimate surface NO2 levels, achieving exceptional performance with a correlation coefficient of 0.975-0.978, a mean absolute bias of 0.99-1.38 ppb, and a root mean square error of 1.47-1.97 ppb. Spatial cross-validation results also indicate strong spatial performance of PCNN-DNN surface NO2 estimates. In addition to its accurate estimates, the PCNN-DNN model consistently generates estimated NO2 grids without any missing values, improving the quality of various applications such as emission reduction strategies and public health studies. Between 2005 and 2019, the 5,478 daily estimated NO2 grids over the CONUS reveal significant reductions in NO2 levels in 14 major urban environments: Washington D.C. (-43%), New York (-45%), Los Angeles (-38%), Chicago (-25%), Boston (-43%), Houston (-34%), Dallas (-40%), Philadelphia (-41%), Phoenix (-38%), Detroit (-20%), Denver (-23%), Atlanta (-0.7%), Cincinnati (-38%), and Pittsburgh (-56%). Furthermore, the study shows that the denser urban regions that in-situ stations are installed in, the higher the difference between in-situ observations and regional-mean NO2 levels. Plain Language Summary Despite the significance of NO2 monitoring stations, their number and spatial coverage is limited, leaving many regions with no NO2 observations. This limitation calls for the development of advanced models to fill the gaps between NO2 stations and create gap-free spatiotemporally-consistent grids of surface NO2 concentrations. Recent advances in deep learning (DL) algorithms have enabled them to fill the gaps in various images, such as those representing the human face, nature, etc. with plausible contents. This technology can also be applied to fill the gaps between surface NO2 stations when the model is trained on datasets such as satellite tropospheric column density of NO2 images, available in almost all regions. In this study, we introduce a novel DL model to construct daily gap-free grids of estimated NO2 levels over the contiguous United States from 2005 to 2019, obtaining exceptional performance with a correlation coefficient of 0.975-0.978. The first phase of the model uses a partial convolutional neural network, trained mainly by satellite images, to fill the gaps between surface NO2 stations, and the second phase utilizes a deep neural network to bias correct the output of the first phase.