A study on red tide detection technique by using multi-layer perceptron

Abstract

This study proposed a red tide detection technique for Cochlodinium polykrikoides(C. polykrikoides) by using Normalized water-leaving radiance from GOCI (: Geostationary Ocean Color Imager) data and 2 stage filtering algorithm based on multi-layer perceptron. We designed the algorithm to classify the type of seawater into 3 classes (Red tide, Clear water, Turbid water). The proposed algorithm has been developed to remove the clear water pixels first and then the turbid water pixels in the satellite images. As a result of the evaluation of the detection accuracy using the verification data, the total accuracy was confirmed to be about 97%. Multi-layer perceptron based algorithm can extract features from input data using their internal structures that is called hidden layer. Nevertheless, it showed similar performance to the logistic regression model using statistical variable selection method and showed higher accuracy than the decision tree model. The results of this study can contribute to reduction of red tide damage through early detection and monitoring of red tide.