Red tide detection using deep learning and high-spatial resolution optical satellite imagery

Abstract

ABSTRACT Red tide is one of the most devastating phenomena that have impacted coastal environments and fishery on a local scale in the worldwide seas. Satellite imagery can provide a synoptic view of the red tides over the wide region. Previous methods to detect the red tides have not sufficiently performed and revealed limitations in the study region. This study developed a red tide detection scheme based on the deep-learning method by using Landsat-8 OLI data during unprecedented explosive red tide events in the southern coastal region of the Korean Peninsula in 2013. To develop and validate the red tide detection of this study, we conducted cruise campaigns to obtain in-situ water sampling for red tide species and its density, chlorophyll-a concentration, suspended particulate matter (SPM), and spectrum data of red tides as a target object and other reference data using a spectroradiometer in the coastal regions from 2013 to 2015. The seawater shows different spectral shape by its components such as red tide, chlorophyll-a concentration, and SPM. Spectral characteristics of the red tides demonstrated bimodal peaks over visible wavelengths regardless of the species of the red tide. Considering such spectral characteristics, the red tide detection algorithm was constructed by the deep-learning method with Landsat-8 OLI level-2 reflectance values and in-situ red tide observations. The validation results of the algorithm as compared with the map of in-situ red tide measurements showed a high probability of detection values greater than 0.7. These works have made it possible to monitor the red tides with high-spatial resolution satellite data and provide a tool to minimize a lot of socioecological impacts from red tide.