Reconstruction of Daily MODIS/Aqua Chlorophyll-a Concentration in Turbid Estuarine Waters Based on Attention U-NET

An attention U-Net was proposed to reconstruct the missing chlorophyll-a concentration (Cchla) data. The is a lightweight full convolution neural network architecture consisting of an enccoder-decoder (i.e., down-sampling and up-sampling). gates (AGs) were integrated into U-Net. Training with AGs could implicitly teach it suppress irrelevant areas highlight salient features in data areas, which would increase sensitivity reconstruction accuracy. uses satellite-derived Cchla anomalies its variance as input, reconstructed fields along their variances outputs. trained applied long-term daily MODIS/Aqua products Pearl River estuary (PRE) adjacent continental shelf area. model performance evaluated by using independent test dataset from both in-situ measurements. results showed that not only had good valid pixels, but also provided more reasonable compared standard without AGs. This study feasible method for task field ocean color, should be helpful producing creditable ecological effects extreme weather conditions such typhoons on upper PRE waters. Based products, footprints studied. surface near typhoons’ track decrease found. composite illustrated increases occurred almost entire area within radius 100 km. time series analysis peak appeared fifth day after typhoon’s passage.