Rolling-SD bloom warnings in Cascade whole-lake experiments (2011–2019) follow the nutrient-driven mean rise; a linear-ramp test rejects a 1.5-fold rise in relative chlorophyll SD

Buelo et al. (2022) report that the rolling-window standard deviation (SD) of daily chlorophyll a, phycocyanin, dissolved oxygen and pH gave early warnings of algal blooms in whole-lake nutrient additions to Peter and Tuesday lakes, with Paul Lake as reference. We used the published daily data (EDI knb-lter-ntl.413.2, 2011-2019) and the original code. The code reproduced the published true-to-false alarm-rate ratios (2.6-8.3) exactly. Then we tested if the SD alarms carry information beyond the rise in mean level that the nutrients caused. On a log10 scale, the chlorophyll SD alarms fell from 56 to 9 alarm days in 332 pre-bloom nutrient days, and the phycocyanin alarms from 55 to 9 in 387 days. Without nutrients, chlorophyll SD increased as mean^1.24 (95% CI 0.90-1.73), and the mean change gave 87% of the pre-bloom log SD ratio between nutrient and reference lakes. A quickest-detection alarm on the chlorophyll mean gave alarms on 54% of pre-bloom nutrient days, with the same 4.3% false-alarm rate as the SD alarm. Surrogates that keep each lake-year's mean path and have constant (relative) noise gave 98% of the observed chlorophyll alarm days and 89-97% for the other variables (500 surrogates; 15-day trend). In synthetic data, a 3-fold rise in relative SD raised the published SD alarm count by only 23-31%, so this alarm responds mainly to the mean. The SD-of-log10 alarm is sensitive to such a rise. The observed count (9) matched the no-rise expectation (median 8) and rejects a 1.5-fold linear rise in relative chlorophyll SD by bloom start (3.5% of replicates as low) and a 2-fold rise for phycocyanin (1%). Conclusion: in these experiments the reported SD warnings are explained by the nutrient-driven rise in mean, and they do not give separate evidence of lost resilience. Limitations: 7 nutrient lake-years, 2 chlorophyll years without nutrients, one reference lake, a linear-ramp alternative model, and published definitions of bloom start and alarm periods.

early warning signals · critical slowing down · algal blooms · cyanobacteria · whole-lake experiment · variance · mean-variance scaling · quickest detection · surrogate data · Cascade lakes · North Temperate Lakes LTER · reproduction