Interpretable ensemble machine learning for assessing algal bloom risk under climate warming

Algal blooms are a major cause of declining lake water quality and are expected to intensify under climate change. Machine-learning approaches have increasingly been used to predict algal blooms; however, most studies emphasise short-term predictive accuracy rather than longer-term bloom risk. In addition, thermal stratification is often treated as a secondary driver, despite its potential importance in a warming climate. Ensemble models are also frequently applied with limited transparency, restricting interpretability and confidence for decision-making.

To address these gaps, we implemented an interpretable ensemble modelling framework to simulate dynamic chlorophyll-a variability using long-term monitoring data from the southern basin of Lake Windermere, UK. The framework integrates multiple commonly used machine-learning models within a transparent stacking structure, calibrated using Bayesian optimisation, and incorporates post-hoc explanation methods to support interpretation of model behaviour and driver importance.

Results indicate that, alongside meteorological and hydrological drivers, temperature-related variables (particularly indicators of thermal stratification) play an important role in controlling chlorophyll-a variability. Scenario simulations were conducted to explore climate sensitivity, including warming experiments, perturbations to thermal stability, and long-term climate projection scenarios. These experiments suggest that warming and increased water-column stability are generally associated with higher chlorophyll-a concentrations and prolonged periods of elevated bloom risk.

Overall, this study presents a transparent and transferable modelling framework for exploring long-term algal bloom risk under climate change, with relevance for lake management and climate adaptation planning.