Abstract. Red tide, a commonly used misnomer for HABs in the Philippines, poses a significant threat to the environment, fisheries, and public health. Since current red tide detection methods, such as in-situ sampling, are mostly reactive and usually result in delayed issuance of advisories, this study developed a model for predicting red tide occurrences in the Philippines using XGBoost and MODIS-derived oceanographic parameters. Five key parameters, namely Chlorophyll-a (chl-a), Sea Surface Temperature (SST), Photosynthetically Available Radiation (PAR), Diffuse Attenuation Coefficient (Kd(490)), and Particulate Backscattering Coefficient (bbp(443)), were extracted from MODIS Aqua 8-day composite products spanning 2003 to 2021. These were integrated with historical data from BFAR, covering the same period, to train a predictive model using the XGBoost algorithm. The final model demonstrated moderate performance as reflected in its accuracy (58%), F1-score (59%), and AUC (61%), with chl-a and Kd(490) as the most influential features based on their SHAP values. Its precision and recall of 58-59% showed its balanced predictive ability across classes, namely, banned and lifted. Model performance across different FMAs and seasons varied due to factors, such as minor variation in parameter values across adjacent FMAs and seasons, missing pixel values of crucial parameters, mismatched parameter values and red tide-linked conditions, and unevenly distributed training data. Among all, the model produced the most reliable and representative results in FMA 7, and the poorest in FMAs 10 and 11.
Arellano et al. (Tue,) studied this question.