A Modified Artifitial Neural Network Ensemble Framework for Drought Estimation

Drought estimation at ungauged sites is a difficult task due to various challenges such as scale and limited availability and information about hydrologic neighborhoods. Ensemble regression has been recently utilized in modeling various hydrologic systems and showed advantage over classical regression approaches to such studies. A challenging task in ensemble modeling is the proper training of the ensemble’s individual learners and the ensemble combiners. In this work, an ensemble framework is proposed to enhance the generalization ability of the sub-ensemble models and its combiner. Information mixtures between the subsamples are introduced. Such measure is dedicated to the ensemble members and ensemble combiners. Controlled homogeneity magnitudes are then stimulated and induced in the proposed model via a two-stage resampling algorithm. Artificial neural networks (ANNs) were used as ensemble members in addition to different ensemble integration plans. The model provided superior results when compared to previous models applied to the case study in this work. The root mean squared error (RMSE) in the testing phase for the drought quantiles improved by 67% - 76%. The bias error (BIAS) also showed 61% - 95% improvement.