Reconstruction of Missing Data in Satellite Images of the Southern North Sea Using a Convolutional Neural Network (Dincae)

A neural network with the architecture of a convolutional au-toencoder is used to reconstruct missing data in satellite images of the Southern North Sea. The technique is applied to a multi-satellite data product of chlorophyll-a and total suspended particulate matter (SPM) concentration (representing 20 years of data). The presence of clouds significantly reduces the extent of the ocean that can be measured by satellite sensors using the visible or infrared spectrum. The accuracy of the reconstruction is assessed using cross-validation (i.e. increasing the actual extent of the cloud coverage). The results of the neural network compare favourably the data withheld for cross-validation.