Abstract
<jats:p>Phytoplankton blooms in lakes and reservoirs are shifting globally in response to climate change, altered nutrient loading, and changes in hydrological patterns, yet most monitoring and research to date has focused only on surface dynamics. Deep chlorophyll maxima (DCMs) can represent a substantial fraction of water-column production and can influence oxygen dynamics, nutrient cycling, and drinking water quality, but long-term, vertically-resolved observations are needed to characterize their interannual variability. To address this gap, we analyzed a decade of ~weekly, high-frequency fluorescence profiles paired with meteorological, physical, and chemical monitoring data in a eutrophic temperate reservoir. We applied random forest models coupled with explainable AI algorithms (SHapley Additive exPlanations, SHAP) to quantify the relative importance and directionality of drivers controlling DCM depth and magnitude. DCMs were a persistent feature for several days to months in every summer across the decade, with significant interannual variation in both depth and magnitude and a directional shift toward shallower DCMs in recent years. Coupled random forest-SHAP analyses revealed that physical drivers (photic zone depth, water level, and thermal stratification metrics) were dominant controls on DCM depth and magnitude. Meteorological and nutrient drivers also contributed meaningfully with non-linear effects on DCM depth and magnitude. These findings demonstrate that DCMs are persistent and respond to a distinct set of drivers, and that combining long-term data and interpretable machine learning can be a useful approach for furthering our understanding of subsurface phytoplankton bloom dynamics.</jats:p>