Machine Learning Approaches for Quantifying Margalefidinium Polykrikoides Bloom from Airborne Hyperspectral Imagery

Abstract

The ichthyotoxic mixotrophic dinoflagellate Margalefidinium polykrikoides is well known for killing fish in aquaculture cages via gill clogging at high cell abundance. In order to establish a countermeasure against M. polykrikoides blooms, it is essential to monitor and quantify the spatial abundance of red tide cells with high accuracy. In this study, machine-learning (ML) approaches were applied to investigate the spectral features of M. polykrikoides cell abundance on the south coast of Korea. Four ML models were developed to retrieve the M. polykrikoides cell abundance in optically complex waters from airborne hyperspectral imagery (HSI). The models included the feed-forward neural network (FFN), support vector machine (SVM), ensemble bagged tree (EBT), and Gaussian process regression (GPR). Paired data of in situ spectra and M. polykrikoides cell abundance were used to train and validate the ML models. The performances of ML models were evaluated using the linear regression value (R2) and root mean squared error (RMSE), considering the predicted cell abundances calculated from HSI and ground truth. The GPR model utilized the validation data and showed the best performance (R2 = 0.57 and RMSE = 659.18 cells mL-1). GPR was also able to estimate the cell abundance with over 3,000 cell mL-1, indicating viable results. Therefore, using GPR is suggested to obtain more accurate red tide cell abundance maps. Local-based ML models for red tide cell abundance retrieval from HSI data can support the monitoring and surveillance of red tide blooms on the south coast of Korea.