Towards a Process Domain‐Sensitive Substrate Habitat Model for Sea Lampreys in Michigan Rivers

Habitat mapping is a common and often useful tool in the ecological management of rivers. The complex nature of fluvial processes, however, makes it difficult to predict the reach‐scale distribution of substrate habitat from landscape‐scale covariates. An option is to identify and partition a data set on boundaries of geomorphic process domains, within which the globally complex relationships between landscape, climate, and instream habitat may potentially be approximated by a simpler model. In this study, we used regression trees as a machine learning method for partitioning and identifying useful strata in a geographically extensive substrate habitat model for larvae of the sea lamprey Petromyzon marinus , an invasive and economically harmful species in the Laurentian Great Lakes. We used field survey data from over 5,000 substrate habitat transects collected in 43 watersheds of the Lower Peninsula of Michigan, and we created a geographic database of geographical information systems‐derived covariates that represent the principal geomorphic influences on substrate habitat. We created three trees in which tree splits delineated (1) spatially contiguous units, (2) noncontiguous units defined by values of the covariates, and (3) both contiguous and noncontiguous units. The adjusted R 2 values of the three trees were 0.30, 0.30, and 0.32, respectively, and all three trees outperformed a single model fitted to the entire data set and a set of models fitted to each watershed individually. The trees identified useful stratifications of Michigan's Lower Peninsula, important geomorphic influences on substrate habitat, and variation in the influence of geomorphic processes on substrate habitat across our study region. Conservation and management applications of our model predictions and tree‐based stratifications include sea lamprey population modeling, habitat survey design, and evaluation of dam removal.