Prediction of biological integrity based on environmental similarity--revealing the scale-dependent link between...

Indices of Biological integrity (IBI) are considered valid indicators of the overall health of a water body because the biological community is an endpoint within natural systems. However, prediction of biological integrity using information from multi-parameter environmental observations is a challenging problem due to the hierarchical organization of the natural environment, the existence of nonlinear inter-dependencies among variables as well as natural stochasticity and measurement noise. We present a method for predicting the Fish Index of Biological Integrity (IBI) using multiple environmental observations at the state-scale in Ohio. Instream (chemical and physical quality) and offstream parameters (regional and local upstream land uses, stream fragmentation, and point source density and intensity) are used for this purpose. The IBI predictions are obtained using the environmental site-similarity concept and following a simple to implement leave-one-out cross validation approach. An IBI prediction for a sampling site is calculated by averaging the observed IBI scores of observations clustered in the most similar branch of a dendrogram--a hierarchical clustering tree of environmental observations--built using the rest of the observations. The standardized Euclidean distance is used to assess dissimilarity between observations. The constructed predictive model was able to explain 61% of the IBI variability statewide. Stream fragmentation and regional land use explained 60% of the variability; the remaining 1% was explained by instream habitat quality. Metrics related to local land use, water quality, and point source density and intensity did not improve the predictive model at the state-scale. The impact of local environmental conditions was evaluated by comparing local characteristics between well- and mispredicted sites. Significant differences in local land use patterns and upstream fragmentation density explained some of the model's over-predictions. Local land use conditions explained some of the model's IBI under-predictions at the state-scale since none of the variables within this group were included in the best final predictive model. Under-predicted sites also had higher levels of downstream fragmentation. The proposed variables ranking and predictive modeling methodology is very well suited for the analysis of hierarchical environments, such as natural fresh water systems, with many cross-correlated environmental variables. It is computationally efficient, can be fully automated, does not make any pre-conceived assumptions on the variables interdependency structure (such as linearity), and it is able to rank variables in a database and generate IBI predictions using only non-parametric easy to implement hierarchical clustering.