EMULATING RAINFALL-RUNOFF-INUNDATION MODEL THROUGH ENSEMBLE LEARNING OF MULTIPLE REGULARIZED REGRESSORS

Abstract

 Recent progress in high-performance computing have enabled meteorological and climate communities to perform large-ensemble weather/climate predictions. To maximize the values of these large ensemble weather/climate prediction data, this study aims to develop a computationally-inexpensive machine that emulates a physics-based rainfall-runoff-inundation model. This emulator predicts maximum inundation depth from the spatial and temporal rainfall data for individual events. We first developed three types of regularized regressors, and then used the Random Forest to conduct ensemble learning of those regressors. This machine structure aims to have two characteristics: preventing over-learning for the underdetermined problem, and stacking multiple weak regressors for the non-linear transformation. There was almost no difference in the predicted accuracy of the maximum inundation depth between three regressors. Regressors tended to underestimate deep maximum inundation depths. The Random Forest significantly improved the prediction accuracy by stacking the weak regressors. In particular, the underestimated inundation depth seen in regularized regressors was greatly improved in the Random Forest.