A predictive model for reach morphology classification in mountain streams using multilayer perceptron methods

This study uses multilayer perceptron (MP) methods to develop classification models for predicting cascade, step-pool, plane bed, and pool-riffle type reach morphologies in mountain streams. Several models were developed with MP and classical linear regression methods on the basis of the following input variables: channel slope (S), sediment size (d₈₄), bankfull depth (h), and bankfull width (w). Data for model calibration and testing were compiled from previous studies in mountain environments. The data were divided into separate calibration (training) and testing (prediction) sets for both the MP and classical linear regression methods; model performance was based on the percentage of accurately predicted reach morphologies using the testing portion of the data. The results indicate that (1) the MP models outperformed the linear regression models for reach morphology classification; (2) relative submergence (h/d₈₄) was useful for classifying step-pool and pool-riffle reaches but performed poorly in discriminating cascade and plane bed type reaches; (3) inclusion of channel slope in models was important for classifying cascade type reaches; and (4) plane bed reaches were the most difficult to classify and delineate from pool-riffle reaches. The two best performing MP models included the input variables (S, h/d₈₄) and (S, h/d₈₄, w). The overall predictive accuracy for classification of reach type for the two models was 81% and 83%, respectively, with predictive accuracies by reach type as follows: cascade, 100%; step-pool, 81%; plane bed, 67%; pool-riffle, 88% (first model) and cascade, 100%; step-pool, 87%; plane bed, 70%; pool-riffle, 90% (second model).