Spatiotemporal predictive modeling for coastal phytoplankton based on the Tapnet model: A case study in Zhejiang,...

Algal blooms have become a global environmental issue, yet research on predicting the spatiotemporal distribution of high-density phytoplankton along the Zhejiang coast remains limited, and existing methods lack highly accurate models applicable to small-sample conditions. Based on historical phytoplankton monitoring data from Zhejiang's coastal waters between 1999 and 2019, this study systematically evaluated and determined the optimal modeling strategy for the Tapnet model. Results: (1) Phytoplankton density exhibits significant interannual fluctuations, with high-density zones primarily distributed in nearshore areas; (2) Results from three feature filtering methods indicate that oceanic environmental features show the strongest correlation with phytoplankton density; (3) Among all approaches, time-series models performed best for predicting high-density phytoplankton distribution along the Zhejiang coast. The optimal Tapnet model achieved an F1-score of 73.58% on high-density samples from 2018 to 2019, outperforming the tree models commonly used in phytoplankton prediction by at least 3.18%. In monthly evaluations, F1-scores exceeded 73.5% for all months except July (0%); (4) Feature filtering strategies struggle to simultaneously enhance prediction performance and computational efficiency; (5) Dimension reduction visualization reveals substantial overlap between high-density and low-density phytoplankton samples in the feature space under the full features dataset. This study identifies an optimal modeling approach for predicting high-density phytoplankton distribution along the Zhejiang coast. It provides a scientific basis for more efficient and accurate monitoring, enabling broader coverage at the same cost and helping mitigate the risks of algal blooms to coastal ecosystems and human health.