NONLINEAR MODELING OF COUPLED WATERSHED PROCESSES USING A DATA-DRIVEN APPROACH

Few studies have attempted to model the coupled and nonlinear anthropogenic, ecological, hydrological, and geochemical processes in a watershed. The primary challenges with a traditional approach are in the joint conceptualization, regularization, and calibration of models for which there typically is disparate and sparse data. We present an alternative nonlinear modeling approach in which a type of unsupervised artificial neural network (ANN) was used to project data associated with 46 subbasins of the Upper Illinois River Basin (28,358 km²) onto a two-dimensional grid called a self-organized map (SOM). The 16,008 data, collected as part of the U.S. Geological Survey, National Water Quality Program, comprised 299 numerical and 49 categorical basin variables from 21 categories (algae, bank stability, climate, demographics, field parameters, fish, geomorphology, habitat, major ions, impervious road, land use, location, macroinvertebrates, nutrients, reach, sediment, segment, site, stream buffer, surficial deposits, and waste water detergents). Pattern analysis in the SOM-based component planes provided information regarding nonlinear correlations and statistically significant variables. Clustering of the SOM topography identified statistically distinct conceptual basin models. Stochastic crossvalidation (using a leave one out strategy) of the ANN model revealed that median variable values were about 95% accurate. A split-sample validation on data from 5 subbasins (not included in the training set) revealed that predictions of statistically significant variables were greater than 78% accurate. Forecasting across urbanizing watershed provided insight on individual changes with respect to integrated processes.