Integrative neural networks models for stream assessment in restoration projects

Abstract

Summary Stream-habitat assessment for evaluation of restoration projects requires the examination of many parameters, both watershed-scale and reach-scale, to incorporate the complex non-linear effects of geomorphic, riparian, watershed and hydrologic factors on aquatic ecosystems. Rapid geomorphic assessment tools used by many jurisdictions to assess natural channel design projects seldom include watershed-level parameters, which have been shown to have a significant effect on benthic habitat in stream systems. In this study, Artificial Neural Network (ANN) models were developed to integrate complex non-linear relationships between the aquatic ecosystem health indices and key watershed-scale and reach-scale parameters. Physical stream parameters, based on QHEI parameters, and watershed characteristics data were collected at 112 sites on 62 stream systems located in Southern Ontario. Benthic data were collected separately and benthic invertebrate summary indices, specifically Hilsenhoff’s Biotic Index (HBI) and Richness, were determined. The ANN models were trained on the randomly selected 3/4 of the dataset of 112 streams in Ontario, Canada and validated on the remaining 1/4. The R 2 values for the developed ANN model predictions were 0.86 for HBI and 0.92 for Richness. Sensitivity analysis of the trained ANN models revealed that Richness was directly proportional to Erosion and Riparian Width and inversely proportional to Floodplain Quality and Substrate parameters. HBI was directly proportional to Velocity Types and Erosion and inversely proportional to Substrate, % Treed and 1:2 Year Flood Flow parameters. The ANN models can be useful tools for watershed managers in stream assessment and restoration projects by allowing consideration of watershed properties in the stream assessment.