Development of Red Tide Detection Algorithm using GOCI Image based on Random Forest

Abstract

The socio-economic damages on the fishery and aquacultural industries caused by the red tide have been increased in Korea. The remote sensing techniques using the ocean color (OC) satellite imagery has been developed in order to observe the red tide. However, the Korean red tide information system (RTIS) is still relying on ship surveillance. It has limitations to cover the whole coastal area as well as take lots of cost and time. This study developed the random forest (RF) based red tide detection model using the Geostationary Ocean Color Imager (GOCI) satellite imagery which has a higher spatio-temporal resolution (i.e., 500 x 500m, hourly). The spectral characteristics, quantitative and qualitative analysis, and spatio-temporal analysis of red tides in the South Sea of Korea during July – August 2018 were examined. The RF model showed promising detection accuracy (R2 = 0.701) than the other three algorithms at high concentrations (over 1,000 cells/mL) quantitatively as well as qualitatively. (i.e., modified red tide index (MRI, R2 = 0.192), red-to-blue ratio (RBR, R2 = 0.683), and spectral shape (SS, R2 = 0.531)). The detection model can provide an accurate red tide alert map in near-realtime as well as contribute to reducing socio-economic damages from the red tides in Korea.