Are more data always better? - Machine learning forecasting of algae based on long-term observations.

Area of Science:

• Environmental Science
• Limnology
• Machine Learning Applications

Background:

• Algal blooms pose risks to ecosystem services and health.
• Effective short-term algae forecasts are crucial for mitigation.
• Machine learning (ML) shows promise for algae forecasting.

Purpose of the Study:

• To determine the necessary volume of training data for reliable ML algae forecasts.
• To assess the impact of training data duration on ML model performance.
• To guide future monitoring strategies and resource allocation for algae bloom prediction.

Main Methods:

• Utilized 30 years of fortnightly measurements of 13 parameters from a UK lake.
• Trained a Random Forest model to forecast chlorophyll-a two weeks in advance.
• Examined the effect of training data duration and feature selection on model performance.

Main Results:

• Random Forest models outperformed benchmarks after four years of training data.
• Model performance improved with over five years of data, but with diminishing returns.
• Using a subset of important features achieved comparable or better performance.
• Reduced sampling frequency negatively impacted forecast performance.

Conclusions:

• Approximately five years of consistent, regular monitoring of key parameters is sufficient for short-term algae forecasting in similar lakes.
• This finding justifies initiating new monitoring programs and utilizing existing datasets.
• Optimized monitoring strategies can enhance the reliability of ML-based algae bloom prediction.