Improving the performance of machine learning models for early warning of harmful algal blooms using an adaptive synthetic sampling method