Reconstruction Of Missing Values In Satellite Remotely Sensed Surface NO₂ Concentration Via Spatiotemporal Feature Learning

Concentration of surface nitrogen dioxide（NO₂）is a crucial air quality indicator which has been routinely measured due to its harmful impact on human health and ecosystems.While numerous ground monitoring stations have been deployed in China for real-time air quality monitoring,the ground NO₂ measurements are only capable of characterizing NO₂ loading at a local scale.Due to vast spatial coverage,satellite remotely sensed column NO₂ concentration data from instruments like Ozone Monitoring Instrument（OMI）onboard Aura satellite,have been widely used to estimate near-surface NO₂ concentration.However,due to the influence of cloud and other factors,extensive missing values in satellite-based NO₂ column concentration data result in incomplete surface NO₂ concentration estimates,thereby challenging regional air quality monitoring and change analysis.To generate seamless near-surface NO₂concentration dataset,here we firstly developed the random forest model to estimate near-surface NO₂ concentration from OMI tropospheric NO₂ column data,and then reconstructed missing values by taking advantage of tensor completion methods.The main contents and findings of this study are summarized as follows:（1）Surface NO₂ concentration prediction model in the Yangtze River Delta region from 2014 to 2020 was established using the random forest method,by taking ground measured NO₂ concentration as the learning target while the OMI-based tropospheric NO₂ column concentration data and meteorological data in conjunction with other auxiliary data used as predictor variables.The validation results indicate a good prediction accuracy,with correlation coefficient（R）of 0.76,root mean square error（RMSE）of 12.66μg/m³,and mean absolute error（MAE）of 8.20μg/m³,respectively.Similar prediction model was then developed by replacing satellite-based NO₂ data with the CAMS reanalysis NO₂ data,enabling to generate gap-free surface NO₂concentration data.The validation results indicate comparable accuracy of this gap-free dataset,with R of 0.74,RMSE of 13.03μg/m³,and MAE of 9.27μg/m³,respectively.Both of these two datasets showed good consistency in spatial distribution and seasonal characteristics with ground measured NO₂ concentration;however,the CAMS-derived NO₂ concentration data were spatially contiguous whereas the OMI-derived NO₂concentration maps suffered from significant data gaps,with an effective temporal coverage of about 120-day per year on average.（2）To deal with extensive data gaps in satellite-derived surface NO₂ concentration maps,the reconstruction low rank tensor completion（RLRTC）method was proposed by taking advantage of tensor completion via singular value decomposition.This method works in principle as an iterative matrix singular value decomposition and reconstruction approach while taking ground measurements as important anchor points to improve local tensor reconstruction accuracy.The experimental results showed that the reconstructed data well depict the observed NO₂ concentration distribution in the Yangtze River Delta in 2020.Compared to the held-out validation data,the reconstructed data showed an annual average R of 0.74,RMSE of 7.75μg/m³,and MAE of 6.11μg/m³,respectively.（3）The reconstruction twist tubal nuclear norm（RT-TNN）method was then developed by integrating two methods of tensor decomposition and tensor product to better reconstruct missing values over regions suffering from large scale data gaps.This method takes spatial correlation of each map into account aiming to better capture the temporal evolution of similar tensor when reconstructing missing values.The validation results showed an annual average R of 0.79,RMSE of 7.21μg/m³,and MAE of 5.72μg/m³,respectively.Overall,two methods of RLRTC and RT-TNN were developed based on matrix and tensor singular value decomposition approaches,allowing for better reconstruction of missing values in satellite-derived surface NO₂ concentration maps.Moreover,the results demonstrated good performance of the two methods,largely benefiting from the consideration of spatiotemporal characteristics of similar tensor data and prior information from ground measurements.These two methods could be used to generate daily gap-free near-surface NO₂ concentration to aid in regional air quality management and exposure assessment studies.