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

Satellite-derived aerosol optical depth (AOD) is negatively impacted by cloud cover and surface reflectivity. As these issues lead to biases, they need to be discarded, which significantly increases the amount of missing data within an image. This paper presents a unique application of the partial convolutional neural network (PCNN) for imputing missing data from the Geostationary Ocean Color Imager (GOCI) by training the PCNN model with the Community Multiscale Air Quality model simulated AOD. The PCNN model outperforms various models and algorithms for imputing GOCI images with a significant amount of missing data (45% of the data set has at least 80% missing pixels) and distance to the nearest known pixel within the GOCI image. Once trained, the model requires significantly less processing time and fewer resources than the other models and methods. The model allows the accurate imputation of remote sensing images within significant amounts of missing data.

Automated summary

The study utilizes a partial convolutional neural network (PCNN) to impute missing data in satellite images, providing more accurate imputation for GOCI images with significant missing data. By training the PCNN model, this approach significantly reduces processing time and resource consumption, outperforming other methods.