Convolutional neural network model for discrimination of harmful algal bloom (HAB) from non-HABs using Sentinel-3 OLCI imagery

Abstract

Harmful algal bloom (HAB) caused by Magalefidinium polykrikoides becomes frequent in Korean coastal waters during the mid-1990s and is now annual events on the southern coast of Korea. HAB often leads to high rates of fish mortality and subsequent economic losses in aquaculture. In addition, non-harmful algal blooms (non-HABs) caused by the dinoflagellate Alexandrium sp., Mesodinium rubrum, and the diatom Skeletonema sp. occur simultaneously in time and space. Because HAB and non-HABs are difficult to discriminate using multi-band satellite data, most previous studies have attempted only detection or qualitative classification with limited data. In contrast, in this current study, we aimed to quantitatively discriminate M. polykrikoides bloom associated HAB from non-HABs around the southern coast of Korea using a convolutional neural network (CNN) model with Sentinel-3 Ocean and Land Colour Instrument (OLCI) imagery with a spatial resolution of 300 m and 16 spectral bands for the first time. To identify the effect of non-HAB patches on the performance of the CNN model, five CNN models were trained with OLCI images as input and ground-truth HAB maps as output data. The appropriate figure-of-merits values (FOMs) with sensitivity of 0.53, precision of 0.92, and F-measure of 0.67 were reasonable when the CNN model trained using the dataset with the highest ratio of non-HABs patches was applied to HAB images. Even when non-HAB images were applied to the models, the CNN model exhibited the lowest error pixel count. Therefore, we confirmed that the CNN model, which can discriminate red tide blooms with subtle differences between the spectrum bands and spatial characteristics, helps solve the complexity and ambiguity in discriminating HAB from non-HABs.

Introduction

Margalefidinium polykrikoides is a red tide species that has caused Harmful algal bloom (HAB) in Korean coastal waters every year since the mid-1990s (Jeong and Kang, 2013, Leeet al., 2013); HAB is generally characterized by a large spatial extent (tens to hundreds of kilometers). Contact with M. polykrikoides bloom water (with cell density of > 1000 cells mL⁻¹) leads to rapid mortality in fish within hours and in shellfish within days (Gobleret al., 2008, Parket al., 2013, Tang and Gobler, 2009, Whyteet al., 2001). This large-scale mortality in finfish causes substantial economic losses. They occur frequently at Goheung, Yeosu, Namhae and Tongyeong on the southern coast of Korea (Fig. 1), mostly during summer, starting from mid-July to early September and lasting until the end of October. Red tide blooms caused by the dinoflagellate Alexandrium sp., Chattonella sp., and Mesodinium rubrum also occur at similar times and locations as M. polykrikoides bloom. Blooms caused by diatoms, such as Skeletonema sp., are like other red tide blooms; there is considerable interspecies competition with dinoflagellate blooms. While high- density blooms caused by these red tide species may adversely affect marine ecosystem due to oxygen depletion, these blooms do not have a harmful effect on the surrounding environment and are thus referred to as non-harmful algal blooms (non-HABs). HAB and non-HABs may occur alone; however, they more often occur in a mixed state. Therefore, for an immediate and efficient response to HAB, it is essential to discriminate them from non-HABs.

The National Institute of Fisheries Science (NIFS), Republic of Korea, has monitored red tides since 1996 (NIFS, 2015). The institute provides daily red tide reports based on field measurements and cell densities that include causative organisms, affected and at-risk areas, and predictions for the development and spread of red tides. The NIFS provides red tide warning system fishermen and aquaculturists based on four levels: red tide emergence attention, red tide attention, red tide alert, and warning lift. For each level, criteria are set for the cell size and toxicity of dinoflagellate and diatom. Red tide warning systems have been issued based on M. polykrikoides blooms due to their harmful properties. In the event of M. polykrikoides blooms, information on these blooms is mixed with other red tide blooms that occur simultaneously. Therefore, information on non-HABs is occasionally included in daily red tide reports. Thus, it is difficult to obtain information on non-HABs alone. In fact, the only way for this information is currently though field sampling at discrete locations from research vessels. Hence, wide or detailed spatial and temporal distributions of red tide blooms cannot be provided. Remote sensing observation based on spectral information has the potential to compensate for these limitations of sparse field measurements.

Discriminating red tide blooms using the absorption properties of phytoplankton (measurable in a laboratory environment) has been attempted. Absorption properties vary depending on the type of pigment and cell size of phytoplankton (Bidigareet al., 1990, Ciottiet al., 2002, Nairet al., 2008, Shanget al., 2014). Thus, measuring the differences in the auxiliary absorption wavelength bands enables phytoplankton species to be identified. Although backscattering coefficient of phytoplankton is much smaller than the absorption coefficient; it is an important optical property that can distinguish red tide species. Cannizzaro et al. (2008) presented a classification technique to differentiate K. brevis blooms from other blooms using a bio-optical dataset consisting of remote-sensing reflectance (R_rs), absorption, backscattering, and chlorophyll (CHL) concentration. Ahn et al. (2009) reported that the shape, absorption, and scattering coefficients vary according to the red tide species. They reported that red tide species had main absorption peaks at wavelengths of 440 and 680 nm but subsidiary peaks at 460, 530, and 590 nm that differed slightly by species. They classified 8 red tide species among 21 species using these optical properties. Tao et al. (2015) compared the absorption and backscattering spectra of Prorocentrum donghaiense and diatoms and identified bloom waters by their low R_rs,555 and high band ratios of R_rs,555/ R_rs,531. Kim et al. (2016) investigated the possibility of optically discriminating red tide blooms by focusing on M. polykrikoides. They generated a dataset of simulated remote sensing reflectance (R_rs) spectra using the Hydrolight software and bio-optical data. They proposed two R_rs band ratios R_rs,555/R_rs,531 and R_rs,488/R_rs,443 for discriminating high-density M. polykrikoides blooms. However, these optical features are difficult to apply to multi-spectral satellite data because of the lack of low spectral resolution and the small number of wavelength bands. In addition, these studies have limitations in the discrimination of red tide species in real complex marine environments due to the difficulty of culturing red tide species, the ambiguity of the distinction between similar species, and the differences between cultivation environments and real marine environments.

Several studies have been conducted to detect red tide blooms using data from ocean color sensors, such as data from MODerate resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imager Radiometer Suite (VIIRS), Geostationary Ocean Color Imager (GOCI), and Ocean and Land Colour Instrument (OLCI) (Huet al., 2015, Izadiet al., 2021, Leeet al., 2020, Lou and Hu, 2014, Qiet al., 2015, Rodríguez-Benitoet al., 2020, Shin et al., 2019, Sonet al., 2012). Terrestrial sensors with high spatial resolution were used for the detection of red tide blooms (Liuet al., 2022, Shin et al., 2019, Shin et al., 2021). These past studies focused on true and false red tide blooms rather than classifying them, even detecting mixed red tide blooms as a single bloom. However, some studies have attempted to differentiate red tide species associated with specific blooms using satellite data. Sathyendranathet al., 2004, Westberryet al., 2005 revealed multi-spectral patterns for the detection of diatom and the cyanobacterium Tricho-desmium spp. based on semi-analytical models. Tao et al. (2015) developed a novel method for discriminating P. donghaiense from diatom blooms in the East China Sea using MODIS data based on the optical properties of blooms. Ghanea et al. (2016) developed a Hybrid Ocean colour Index (HOCI) to distinguish Trichodesmium erythtraeum and M. polykrikoides from other red tide species in the Persian Gulf using MODIS data. Shin et al. (2017) proposed that red tide species which occur frequently in Korean coastal waters could be divided into two groups according to their CHL contents; they developed a system for red tide surveillance using GOCI images. Ghatkar et al. (2019) classified three major algal blooms, including Trichodesmium erythraeum, Noctiluca scintillans, and M. polykrikoides blooms, from remote sensing data using an extreme gradient-boosted decision tree model. Feng et al. (2020) developed a method based on backscattering for discriminating summer blooms of harmful raphidophyte (Chattonella sp.) and the diatom (Skeletonema sp.) using MODIS images in the Ariake Sea, Japan.

Until now, due to the difficulties of image collections related to various red tide blooms and the ambiguity of spectral distinction between red tide blooms in the satellite imagery, studies have only attempted to discriminate red tide blooms qualitatively using satellite data. Therefore, methods based on existing red tide indexes have limitations in discriminating red tide blooms. To overcome these limitations, deep learning approaches can be effective. Recently, convolutional neural network (CNN) model, which can resolve a problem of non-linear relationship between spectrum of red tide blooms, have been used for red tide detection. CNN model is an essential tool for deep learning and is particularly suited for image recognition (Yamashita et al., 2018). Kim et al. (2019) proposed the automatic pixel-based detection of three red tide groups using a deep CNN model, U-Net, from GOCI images of the coast of the Korean Peninsula. The predicted red tide bloom maps showed considerabe matching distribution of the three groups to the ground truths. Shin et al. (2021) developed a U-Net deep learning model for detecting M. polykrikoides blooms off the southern coast of Korea from PlanetScope imagery. The predicted map derived from U-Net provided reasonable red tide patterns for all water areas. However, no studies have attempted to discriminate M. polykrikoides blooms from other red tide blooms using a CNN model.

In this study, we aimed to quantitatively discriminate M. polykrikoides bloom from various red tide blooms along the southern coast of Korea. For this purpose, we visually inspected various red tide blooms using Sentinel-3 OLCI with a spatial resolution of 300 m and 16 spectral bands and investigated the OLCI-based spectral characteristics of HAB and non-HABs. We trained and tested a simple CNN model to discriminate between HAB and non-HABs. Finally, we qualitatively and quantitatively validated the CNN models using OLCI images.