Prediction of Hydrological Classes of Streams Based on Watershed Attributes

The implementation and results of three methods for predicting hydrological classes of streams from watershed attributes is presented. Classifications that group streams into hydrologically different classes are often developed based on streamflow data for various applications (for e.g., assessing hydrologic alteration, regionalization methodology, ecological studies etc). A challenge in using such classification schemes is that in most of the cases streams are ungauged necessitating the need for prediction of hydrologic classes. From our previous study, we had hydrological classification schemes based on streamflow regimes for 543 Hydro Climatic Data Network sites in the Western USA. In the present work, we used the classification schemes from our previous study to demonstrate three methods; a) Linear Discriminant Model (LDM), b) Classification and Regression Trees (CART), and, c) Random Forests (RF), to predict the hydrologic class of an ungauged watershed from watershed attributes. Watershed attributes consists of climate, geomorphology, geology, and soils variables. The classification error and predictive uncertainty from these models was estimated using a Monte-Carlo approach. We found that the RF model gave a slightly better prediction of stream class than the other models evaluated (% correct predictions = 67 for RF, 57 for CART and 63 for LDM). The predictive uncertainty was similar for all the three methods (standard deviation of % correct predictions= 5.7for RF, 5.9 for CART and 6.3 for LDM). In all the three methods, climate related watershed attributes were the most dominant in discriminating the hydrological classes suggesting that at the scale of Western USA, climate is the most important discriminator of hydrology.