Early Prediction of Margalefidinium polykrikoides Bloom Using a LSTM Neural Network Model in the South Sea of Korea

Abstract

Shin, J.; Kim, S.M.; Son, Y.B.; Kim, K., Ryu, J.-H., 2019. Early prediction of Margalefidinium polykrikoides bloom using a LSTM neural network model in the South Sea of Korea. In: Jung, H.-S.; Lee, S.; Ryu, J.-H., and Cui, T. (eds.), Advances in Remote Sensing and Geoscience Information Systems of Coastal Environments. Journal of Coastal Research, Special Issue No. 90, pp. 236-242. Coconut Creek (Florida), ISSN 0749-0208.Harmful algal blooms (HABs) have been occurring within the South Sea of Korea (SSK) for decades, causing significant ecological impacts and economic problems for many fish farms concentrated in coastal areas. The occurrence of HABs is related to various factors, including meteorological and physical factors, making it difficult to predict the timing of their occurrence. However, it is essential to make a preliminary forecast of HAB occurrence through analysis of such factors to minimize the damage. In this study, a deep neural network model of long short-term memory (LSTM) that can predict the occurrence time of Margalefidinium polykrikoides blooms is presented and evaluated. Satellite data were used to extract sea surface temperature (SST) and photosynthetically available radiation (PAR), which are environmental factors known to be related to HABs occurrence. M. polykrikoides blooms that have occurred in the past 21 years (1998–2018) have been shown to be initiated when SST reaches around 25 °C within the summer season. The prediction performance of LSTM-based neural network was evaluated using test data from 2017 and 2018, and the root-mean-squared-error (RMSE) and prediction accuracy were determined. Prediction for 2017 matched well with the test data; however, for 2018, the network predicted that a HAB would occur between July 19 and Sep. 03, with an RMSE of 0.233 and accuracy of 94.8 %, but the actual occurrence of HAB was between July 24 and Aug. 20. This study shows that the trained LSTM-based network would be useful for early prediction of the future red tide blooms in SSK.