Use of deep learning and a partial convolutional neural network to gap-fill a long term time series of NO2 columns from satellite impacted by cloud

Satellite monitoring plays a significant role in monitoring nitrogen dioxide (NO2) concentrations in the atmospheric column, but it is often affected by clouds and ice and snow surface. This leads to much missing data. Deep learning with Partial Convolutional Neural Network (PCNN) is adept at handling incomplete or missing data in image processing by focusing only on the known pixels during convolution, thus making approach ideal for tasks such as image restoration, denoising, and enhancing resolution. It is therefore important to reduce such data gaps.. Under cloudy skies, ground-level NO2 often tends to be higher. Clouds are typically associated with low pressure and increased wind speeds in mid-latitudes, leading to enhanced dispersion of pollutant. However, low cloud often occurs during periods of high pressure when boundary layer heights are lower and air pollutants are trapped closer to the ground. Additionally, clouds intensify the Surface Sensible Heat Flux, contributing to the urban heat island effect and potentially increasing NO2 concentrations. On the other hand, clouds decrease Surface Net Solar Radiation, which might mitigate NO2 photolysis. It is therefore likely that NO2 concentrations close to the surface during cloudy conditions will not necessarily be well represented by satellite derived NO2 columns in clear sky conditions.. It becomes necessary to recalibrate satellite-derived data to reflect actual meteorological conditions. In this work we separate out ground-level data from an urban network across Paris, France, into two categories: those with contemporaneous TROPOMI and those without. Each category is then analyzed with the weather conditions at that time. This analysis helps estimate the variance in NO2 concentrations due to cloud presence. Subsequently, the determined percentage difference, indicative of the cloud cover's impact, is applied to the NO2 estimates provided by the PCNN model. This adjustment not only strengthens the data's coverage but also its reliability, reducing the biases in the original satellite data resulting from clear sky viewing only and are therefore a closer representation of the urban atmospheric pollution. This approach, combining technical precision with contextual sensitivity, improves the use of satellite data as a tool for understanding and interpreting urban pollution.