Application of a Partial Convolutional Neural Network for Estimating Geostationary Aerosol Optical Depth Data

Abstract

A large fraction of secondary aerosols is produced from the condensation of precursor gases or nucleation via cloud processes (Ervens et al., 2011). Particulate matter (PM), a major air pollutant worldwide (Koulouri et al., 2008;Mukherjee & Agrawal, 2017), comes in two aerodynamic diameters of fine particles: less than 10 µm (PM 10 ) and less than 2.5 µm (PM 2.5 ) (United States Environmental Protection Agency [USEPA, 2020]). As PM is associated with respiratory and cardiovascular diseases and mortalities (Brunekreef & Holgate, 2002), acquiring accurate estimates of PM is important to assess its impact on human health. Estimates have shown that globally over 2 million deaths resulting from damage to the respiratory system per year are associated with PM pollution (K. H. Kim et al., 2015;Shah et al., 2013). A common product of satellites for estimating PM levels (Ghahremanloo et al., 2021) is the aerosol optical depth (AOD). Although researchers have devoted a great deal of effort to develop systems that improve PM forecasting, they have been limited by the quality and amount of data (Ghahremanloo et al., 2020) and the performance of chemical transport models (Park et al., 2011;Pouyaei et al., 2020). To address the limitations of models and improve the accuracy of models for atmospheric chemistry and weather forecasting, they have incorporated both in situ measurements and

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

“…Previous studies (Di et al., 2016; Ghahremanloo et al., 2022; Hu et al., 2017) have used simple approaches such as the inverse distance weighting (IDW) interpolation approach to create the interpolated grid of air pollutants. However, the application of DepthWise PCNN can produce more accurate results since DL models can better capture the spatial pattern of the pollutants (Lops et al., 2021). The second phase of the PCNN‐DNN model uses a DNN model comprised of one input layer, four hidden layers with 70 neurons each, and one output layer to estimate surface NO 2 levels over the CONUS from 2005 to 2019.…”

“…Although researchers (Bugeau et al, 2010;Y. Li et al, 2017;Lops et al, 2021;Yu et al, 2011) have used various inpainting approaches to impute missing remote sensing data, most of these approaches have serious limitations, imputing images with a high frequency of missing data (Lops et al, 2021). To address this limitation, the partial convolutions approach, introduced by G. Liu et al (2018) imputes images with a high frequency of missing data.…”

“…To address this limitation, the partial convolutions approach, introduced by G. Liu et al (2018) imputes images with a high frequency of missing data. In this regard, Lops et al (2021) leveraged the partial convolutional neural network (PCNN) to impute missing pixels in the Geostationary Ocean Color Imager (GOCI) aerosol optical depth (AOD) product. They reported promising performance of their model at imputing AOD images with a significant number of missing pixels.…”

A global assessment of the mixed layer in coastal sediments and implications for carbon storage

“…Neural networks have emerged as a powerful tool to resolve complex spatial and temporal patterns that are common in large datasets in the geosciences 17,18 . Some examples of neural networks applications in the earth sciences include: hydrology, including flooding forecasts, and water quality modeling 19,20 ; geophysics/ geomorphology, including earthquake predictions, and simulating land-use change 21,22 ; and atmospheric sciences, by modeling cloud formation and temperatures 23,24 . Recently, machine learning (e.g., K-nearest neighbor, random forest and neural networks) has been applied to resolve important questions in oceanography, including the global distribution of seafloor total organic carbon, benthic properties, sediment porosity/density and sediment accumulation rates [25][26][27][28][29][30] .…”

Fast and operational gap filling in satellite-derived aerosol optical depths using statistical techniques

“…They have imputed AODs derived from MODIS observations by combining two techniques: random forest (a popular nonparametric machine learning algorithm) and lattice Kriging (a multiresolution Gaussian process model). 7 To fill gaps in geostationary ocean color imager (GOCI) AODs using CTM simulations, Lops et al 8 proposed a novel deep learning technique, the partial convolutional neural network, which showed acceptable performance. However, including CTM simulations in the training process, as in the aforementioned studies, can increase computational costs, especially when they use the model to achieve high spatial resolution.…”