River Stage Forecasting Using Multiple Additive Regression Trees

Abstract

Accurate real-time forecasts of river stages can serve as a reference for flood evacuation to minimize losses and casualties. Machine learning has been widely used for river stage forecasting because of its simple modeling and quick computation. However, many machine learning models have drawbacks such as excessive learning time, difficult evaluation of input variables, and lack of explanatory capacity, which limit their performance as practical tools. To overcome these drawbacks, this study employs multiple additive regression trees (MART) for river stage forecasting. Three MART models are proposed, namely the original MART model, the real-time MART model, and the naive MART model, with different considerations of model training and error correction. Model training and testing were conducted based on the rainfall and river stage data for 16 typhoon events between 2005 and 2009 in the Bazhang River Basin in Taiwan. In the training process, variables are automatically selected by the MART models which reasonably describes the mechanism of flood transportation. The testing results show that all three models can reasonably forecast the river stages with a three-hour lead-time. Compared with the original MART, the real-time MART performs better in describing overall river stage variations, whereas the naive MART is more accurate in the prediction of peak river stages. The proposed MART models are efficient and accurate and can thus serve as practical tools for flash flood early warning.