In eutrophic coastal waters, harmful algal blooms (HAB) often occur and pose significant challenges to fisheries management and food and water security. The onset of a HAB is notoriously difficult to predict. Despite decades of research on HAB early warning systems, the field investigation of algal bloom forecast models has received scant attention. Traditional approaches of red tide monitoring and fisheries management rely on field sampling and laboratory analysis of chlorophyll-a concentration (Chl-a) - an indicator of algal biomass - and manual cell counting and species identification, which are resource-intensive and time-consuming.

With the increasing availability of real-time water quality sensors and high-frequency algal imaging tools, the development of HAB early warning systems...[ Read more ]

In eutrophic coastal waters, harmful algal blooms (HAB) often occur and pose significant challenges to fisheries management and food and water security. The onset of a HAB is notoriously difficult to predict. Despite decades of research on HAB early warning systems, the field investigation of algal bloom forecast models has received scant attention. Traditional approaches of red tide monitoring and fisheries management rely on field sampling and laboratory analysis of chlorophyll-a concentration (Chl-a) - an indicator of algal biomass - and manual cell counting and species identification, which are resource-intensive and time-consuming.

With the increasing availability of real-time water quality sensors and high-frequency algal imaging tools, the development of HAB early warning systems has become a practical possibility. In this study, two major aspects of a HAB early warning system are explored: (1) daily forecast of algal bloom risk based on the prediction of sea surface temperature and vertical density gradient using in-situ real-time (10-min sampling interval) water quality data; (2) automatic identification of target HAB species using machine learning techniques from high-frequency images (up to 40,000/hr) monitored by a submerged Imaging FlowCytobot (IFCB) at a marine fish farm. The two aspects can be integrated to significantly enhance the ability to observe and monitor HAB events in coastal waters.

The daily algal bloom risk forecast system is developed based on: (i) a vertical stability theory verified against 191 past algal bloom events; and (ii) a data-driven artificial neural network (ANN) model that assimilates high-frequency data to predict sea surface temperature (SST), vertical temperature and salinity differential daily. The model does not rely on past chlorophyll measurements and has been validated against extensive field data. Operational forecasts are illustrated for representative algal bloom events at two representative marine fish farms in Hong Kong waters with different hydrographic conditions. One is Yim Tin Tsai in the Tolo Habour in Eastern Waters that are relatively sheltered from Pearl River input. The other is Lo Tik Wan near the Lamma Island at the Southern Waters that are strongly affected by Pearl River flow.

In the automatic species identification system, the IFCB uses hydrodynamic focusing to acquire algae images, which can be auto-classified using machine learning. A 3D computational fluid dynamic model has been developed to study the hydrodynamic focusing flow in IFCB and verify its performance. An explainable supervised machine learning technique has been successfully developed based on image data collected during two years of submersible deployment. The algal species classifier is trained by presenting a wide range of extracted image features to a random forest algorithm. A recursive feature elimination technique identifies an optimized set of 25 features. The random forest (RF) classifier can identify 15 target HAB classes with an overall out-of-bag accuracy of 94.2%, with individual F1 scores ranging from 0.8 to 1.0. The classifier performs equally well as a Convolution Neural Network (CNN) developed using transfer learning techniques (all F1 scores > 0.8). Based on the classifier, automated real-time species identification and cell counting protocol have been developed, with a response time of 10 minutes after data collection.