Predicting river bed substrate cover proportions across New Zealand

Predictions of river bed substrate cover are required for various purposes including delineating management zones, linking with ecological status and assessing river rehabilitation options. Three contrasting methods were tested for predicting the proportion of river bed covered by seven different substrate categories: generalised linear models (GLMs), machine learning regression models (random forest), and a summed normal distribution model (SND) which incorporates distribution of predictors and substrate covers throughout the modelling framework. Various predictors representing climate, geomorphology, land cover and geology were derived from existing environmental databases to generate predictive models. Model performance was assessed through a cross-validated comparison with substrate samples collected from 229 river sites distributed across New Zealand. Model performance for 10-fold cross-validated predictions showed that the SND model performed best in predicting the proportions of riverbed covered by bedrock, boulder, cobble and fine gravel categories. Random forest models performed best in predicting coarse gravel, sand and mud plus vegetation proportions. Therefore, combined random forest and SND methods were used for estimating substrate cover proportions at unsampled sites across New Zealand. Texture analysis of predicted substrate cover consistently showed downstream fining of sediment size. The national predictions of substrate cover proportions are key descriptors that can be linked with a wide range of national scale applications for ecological assessment of New Zealand Rivers. The techniques developed and tested are applicable to other locations but it is notable that relatively poor performance in regional cross-validation tests shows that transferability of substrate models to locations with no calibration data is challenging.