Predicting Distribution of Juvenile and Adult Mussels in Pool 18 of the Upper Mississippi River Using Random Forest Models

We analyzed data from a quantitative survey of native mussels that was conducted in a 42-km impounded reach of the Upper Mississippi River (Navigation Pool 18) using a systematic design (n=367 sites).   For each sampling site, we estimated simple physical and complex hydraulic variables that have been shown to be useful descriptors of mussel habitat in the Upper Mississippi River.   Presence-absence of juvenile and adult mussels were analyzed with random forest models.  This ensemble learning method aggregated classification tree submodels (N=1200) based on random selection of predictor variables and data.  To reduce the effect of prevalence on predictions, models were constructed using down-sampled data to balance sample sizes of presence and absence.   Out-of-bag samples were used to estimate generalization error.   Receiver operating characteristic curves indicated useful models were constructed for both adult and juvenile mussels.  However, the model for adult mussels performed considerably better (Area under the Curve, AUC=0.81; overall error rate=22%) than the juvenile mussel model (AUC=0.72, overall error rate=33%) indicating greater predictability for adults.   Models depended on complex hydraulic variables such as relative substrate stability, boundary Reynolds number, and Froude number, but pool thirds (a measure of longitudinal position in the reach) and bathymetry were also important.