Machine learning can refine models of environmental suitability and ecological associations for invasive quagga mussels

Abstract Dreissenid mussels are among the most prolific aquatic invasive species globally, damaging water industry infrastructure, altering freshwater ecosystem functioning, and costing the global economy millions of US dollars annually. Extensive research has been conducted to predict, prevent, and understand the impacts of dreissenid spread for zebra mussels ( Dreissena polymorpha ). However, similar efforts for quagga mussels ( D. rostriformis bugensis ), the more prevalent dreissenid in the western United States, have lagged. To better characterize quagga habitat suitability and trophic relationships across six hydrologically connected Arizona waterbodies, we collected water quality and plankton community data at 20 sampling stations from 2021 to 2023. Four waterbodies had established quagga populations (hereafter “established”), while the remaining two did not (hereafter “negative”), despite suspected opportunities for introduction. Using data reduction techniques and an advanced machine learning (ML) classification algorithm, gradient boosted machine, we identified environmental and ecological conditions that differentiated established from negative stations. Notably, parameters considered crucial to dreissenid invasion (e.g., calcium, alkalinity, temperature, dissolved oxygen) were not among the most important variables to classification, as all examined waterbodies exhibited quagga‐suitable ranges. Rather, in this system, the established class was characterized by conditions linked to dreissenid osmoregulation (e.g., higher total dissolved solids and potassium) and indicators of primary productivity and trophic state (e.g., higher chlorophyll a, total phosphorous, and total nitrogen). Further, established stations had lower zooplankton abundances of dreissenid competitors (e.g., Bosmina longirostris , cyclopoid copepodids) and prey (e.g., Keratella sp., Polyarthra spp.), perhaps resulting from food competition and consumption, respectively. Notably, negative waterbodies, Bartlett Reservoir (Bartlett) and Theodore Roosevelt Lake (Roosevelt), exhibited different biotic and abiotic conditions from each other. Stations in both showed indications of lower trophic states, yet Roosevelt exhibited higher densities of quagga‐competitor zooplankton, while Bartlett displayed poorer conditions for quagga osmoregulation. Our study illustrates how ML can identify quagga environmental and ecological relationships within established waterbodies and demonstrates that negative waterbodies with suitable calcium concentrations may still have differing invasion risks. Therefore, ML can assist managers in prioritizing risk reduction efforts to waterbodies with less robust invasion inhibiting factors, rather than those with stronger abiotic defenses.