Development of Artificial Intelligence Approach to Nowcasting and Forecasting Vibrio Prevalence in Coastal Waters

Vibrio parahaemolyticus (V.p) is an epidemiologically significant pathogen that poses high risks to the human health and shellfish industry, calling for predictive models for management interventions. This study presents an Artificial Intelligence(AI)-based approach to predicting and reducing the risks. The AI-based approach involves the identification of environmental indicators and their optimum variation ranges favoring V.p prevalence, the development of nowcasting and forecasting models for predicting V.p prevalence, and the creation of remote sensing algorithms for mapping concentrations of V.p and its environmental indicators by synergistically combining the Deep Neural Network (DNN) modeling technique, Genetic Programming (GP) method, R statistical computing environment, and satellite remote sensing imagery data. In terms of environmental indicators and their optimum variation ranges it was found that the importance of environmental indicators (and their optimum ranges) to the prevalence of Vibrio parahaemolyticus can be ranked from the highest to lowest as Sea Surface Temperature (SST: 25.67 ± 2°C), salinity (27.87 ± 3 ppt), water level, pH (7.96 ± 0.1), chlorophyll a, and turbidity, respectively. In terms of nowcasting models, three stacking Deep Artificial Neural Network (DNN) models with Genetic Programming (GP) functions of SST, salinity, and water level as model input variables were presented. The stacking GP-DNN nowcasting models are best suited when in-situ data are available. Another nowcasting model was created for mapping V.p concentrations by using high-resolution satellite remote sensing data from advanced Sentinel-3 OLCI and SLSTR instruments in case the in-situ data are not available. In terms of forecasting models, a Random Forest-based forecasting system consisting of four models with differing lead times of 1 – 4 days was constructed for predicting the V.p abundance in oysters, providing both the sufficient lead time for emergency preparedness and intervention and the high predictive performance. In order to overcome the in-situ data scarcity issue and map vibrio infection risks in oyster harvest areas, remote sensing algorithms were created for major environmental indictors (particularly SST, salinity and chlorophyll-a) of V.p, making it possible to nowcast and forecast V.p infections even if in-situ data are not available. The findings from this dissertation are of both theoretical and practical significance. In terms of theoretical significance, the optimum ranges of environmental indicators clarify the existence of both positive and negative correlations reported in the literature and thereby resolve the long-standing dispute on whether the relationship between the Vibrio prevalence and environmental conditions should be positive or negative. In terms of practical significance, the remote sensing-assisted nowcasting and forecasting models enable public health agencies and oyster harvesters to focus more on preventing V.p infections, rather than relying on reacting to problems after they have occurred, greatly reducing the risk of V.p infection to human health and the risk of economic loss to the seafood industry