Spatial Information-Based Back-Propagation Neural Network Modeling for Outflow Estimation of Ungauged Catchment

Rainfall-Runoff Models (RRMs) are standard tools for hydrologists to organize water resource planning and management. With the development of space-information technology and computing science in recent years, RRMs have improved to better simulate the rainfall-runoff process along with the spatial variation of simulated catchment, with the aim of precisely articulating the underlying relationship between input and output information. The Artificial Neural Network (ANN) model is suitable because it has various mathematical compositions capable of simulating a nonlinear structural system to establish the flow discharge in a catchment. For this reason, it has been successfully applied in RRMs modeling. In this study, the rainfall-runoff behavior of the gauge stations will be replicated in an ANN model. A Back-Propagation Neural network (BPN) was adopted to estimate an ungauged region's outflow considering the temporal distribution of rainfall-runoff and the spatial distribution of the watershed environment. The nonlinear relationship among the physiographic factors, precipitation and outflow of the specific catchment was established to estimate the outflow of the sub-catchment where no flow gauge had been provided. The hazard preventive hourly model was mainly considered in this paper, and the model was also compared with the Instantaneous Unit Hydrograph (IUH) and Area Ratio Method (ARM) approaches. In validation tests in the middle and large catchments in Taiwan, the model performed well in providing forecast results. From the research work, it is revealed that the relevant spatial information can be obtained easily and precisely, which will help future studies on more related dimensions.