Development of a Predictive Model for Early Detection of Red Tide Occurrence Using Change-point Analysis and Logistic Regression in the Ariake Sea

Abstract

Using continuous winter water-quality data from the Ariake Sea, the early detection of red-tide occurrence was examined by time-series statistical analysis. Applying variance change detection to chlorophyll a (CHL) revealed three distinct phases: normal, transition, and divergence. Correlation and partial correlation analyses revealed that, during the transition period, the relationships with water depth and salinity reversed, suggesting phase changes in water quality and the ecosystem driven by variations in tidal and freshwater inflow. Considering pH as a precursor indicator rather than a causal factor, a logistic regression model was constructed using the mean and variance of pH as features, with CHL≧10 μg L^-1 as the binary response variable for red-tide occurrence. This model demonstrated high discriminative performance: AUC = 0.979, sensitivity = 0.960, specificity = 0.917, and overall accuracy = 0.932. Furthermore, the probability of red-tide occurrence increased stepwise during the transition period following the CHL variance change point, with peak values observed approximately 2–5 days before the transition to the divergence phase. These results indicate that the predictive model based on change-point analysis and pH time-series features has the potential to provide operationally useful early-warning information several days before red-tide expansion.