The effect of collinearity between observed and model derived training variables on estuarine algal species distribution models

Forecasts of organism distributions in time and space are needed to mitigate risks associated with changes in environmental conditions. These forecasts are often generated using correlative species distribution models (SDMs) that relate environmental variables to species presence or abundance. Biological complexity makes the construction of SDMs challenging because the collinearity between the environmental variables used to train the SDM may increase model parameter uncertainty. To analyze the effect of collinearity on SDMs, we (1) train SDMs for seven estuarine algal species commonly observed in the Chesapeake Bay (U.S.A.) using different levels of collinearity in the training information, (2) identify the environmental predictors, and (3) study their association with species presence using two statistical techniques (generalized linear models and regression trees). The novelty of our contribution is that our analysis uses both environmental in situ observations and environmental information generated by a mechanistic model. The environmental variables show strong collinearities in both the in situ observations (32 out of the total of 165 correlations) and mechanistic model output (12 out of the total of 120 correlations). To determine how collinearity between these variables affect our SDM results, we remove environmental variables that surpass a specific correlation threshold. We find that using these two different types of training information (i.e., observed vs. modeled) affects (1) the optimal set of predictors, (2) the associations between environmental variables and algal presence, and (3) the model’s predictive skill. Water temperature is generally selected as an important predictor. Strong positive or negative associations between environmental variables and algal presence are not substantially impacted by the type of training information used. Although removing collinearities may result in the detection of new important predictors, it may also result in a slight decrease ( ∼ 5 %) of the SDM prediction skill, depending on the species of interest and type of training information. Our findings suggest that the main environmental predictors rely on both the species characteristics and training information type used in SDM construction, and highlight the challenge of interpreting the associations between environmental conditions and species presence predicted by these SDMs. These insights help us to better understand the environmental conditions of importance to these algal species and hence optimize monitoring efforts by revealing which in situ observations are vital to accurately forecast blooms of these estuarine algae. • We use field data and model outputs to train SDMs for seven estuarine algal species. • Reducing collinearity reveals new environmental predictors that were hidden before. • Training the SDM using in situ vs. model outputs also influences this optimal set. • Temperature and salinity are the main predictors, regardless of species or training data. • Some algal species share similar environments, making SDM predictions challenging.