Deep Convolutional Neural Network for Spatial Red Tide Detection

Abstract

It is important to detect promptly red tide blooms in a vast ocean for protecting marine ecosystem. In this study, we proposed a deep convolutional neural network (CNN) for automatic spatial red tide detection. In order to train the CNN, we constructed training data with GOCI images between 2011 and 2018 and other environmental factors such as sea surface temperature (SST) and photosynthetically active radiation (PAR). The CNN can learn the spectral features and the correlation related to the occurrence of red tides from GOCI spectral images and the environmental factors. The spectral GOCI images are taken hourly over Northeast Asian region at 8 times a day with spatial resolution of 500 m. The daily SST and PAR data come from MODIS and GHRSST. Three spatial data have different spatial resolution, thus we did pre-processing to match the resolution over the interested region. As ground truth indicating where the red tides were occurred, we considered a red tide index map generated by a decision tree and real measurements provided from National Fisheries Research & Development Institute. The trained CNN gave well-matched red tide index maps to ground truth.