Comment on "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)" by Recknagel et al. (Water Research 287 2025 124,514).

Recknagel et al. (2025) present a timely study leveraging high-frequency in-situ data and three fundamentally different machine learning algorithms to forecast cyanobacterial blooms in Lake Müggelsee at a 5-day horizon. However, we note that four methodological choices may overstate actionable early-warning skill if not accompanied by robustness checks: (i) cross-year preprocessing and interpolation of weekly biovolume to hourly/daily resolution, which can introduce temporal leakage and smooth bloom dynamics; (ii) evaluation emphasizing R2/RMSE rather than threshold-exceedance skill and lead-time distributions at the 4 mm3/L hazard level, with limited uncertainty communication; (iii) strong reliance on phycocyanin (PHYCO), which is operationally reasonable but warrants PHYCO-persistence baselines and no-PHYCO ablations to quantify incremental forecasting value and sensor dependence; and (iv) mixed aggregation choices and single-station drivers paired with multi-station/depth-integrated targets, which can bias learned relationships. A practical set of leakage-robust validation practices, sensitivity analyses, event-based metrics, uncertainty reporting, and strong baselines could further strengthen the study's operational credibility without changing its overall contribution as a pragmatic demonstration.