Early warning of harmful cyanobacteria blooms based on high frequency in situ monitoring and intelligible machine learning modelling: The case study of Lake Müggelsee (Germany)

Driven by eutrophication and global warming, the occurrence and frequency of harmful cyanobacteria blooms (CyanoHABs) are increasing worldwide, posing a serious threat to human health and biodiversity. Early warning enables precautional control measures of CyanoHABs within water bodies and in water works, and it becomes operational with high frequency in situ data (HFISD) of water quality and forecasting models by machine learning (ML). However, the acceptance of early warning systems by end-users relies significantly on the interpretability and generalizability of underlying models, and their operability. To achieve acceptance, following principles of intelligible modelling of CyanoHABs by ML algorithms are suggested: (1) HFISD should be of reasonable length (≥ 5 years) and frequency (≤ 1 day) to reflect typical patterns of CyanoHABs. (2) Driver variables should be causally related to cyanobacteria properties. State variables for cyanobacteria require in situ measured data to be comparable with 'high hazard level'. (3) ML algorithms must be suitable for time-series modelling. (4) Forecasting models must be generalizable, interpretable and operable as prerequisite for real-time early warning of CyanoHABs. Intelligible modelling of CyanoHABs is demonstrated by the case study of Lake Müggelsee (Germany) where the performance of the three fundamentally different algorithms XGBoost (decision trees), LSTM with attention (deep learning neural networks) and HEA (causally inferred rules) is assessed. Based on 11 years of hourly and daily HFISD, the models were designed for 5-day-ahead forecasts of total cyanobacteria driven by water temperature (WT), turbidity (TURB), pH and phycocyanin (PHYCO). When daily data were used, HEA and LSTM models predicted CyanoHAB events exceeding the 'high hazard level' in training data while only XGBoost and HEA models achieved satisfying results for unseen data. Based on hourly data, all three models predicted CyanoHAB events exceeding the 'high hazard level' in the training data whereby only XGBoost and HEA performed satisfyingly for unseen data. Forecasting results for unseen data testified good generalizability of rule-based models by XGBoost and HEA. XGBoost and LSTM didn't represent models explicitly. By contrast, HEA gained high interpretability by logic IF-THEN-ELSE-rules that explained the inner working of the models. Post-hoc data interpretation revealed that driver importance of all three models ranked PHYCO highest but only XGBoost and HEA ranked WT second highest that was ranked lowest by LSTM. Using daily HFISD enabled slightly better forecasting results compared to hourly data that may have implications for the monitoring and modelling process. Overall, intelligible ML modelling of HFISD proved to be prerequisite for the development of evidence-based and actionable real-time early warning systems for hazardous ecological events such as CyanoHABs. It became evident that rule-based models developed by XGBoost and HEA generalized better than models by LSTM and that logic rule models by HEA achieved highest interpretability and operability.