Rainfall-Runoff Modeling Using Support Vector Machine in Snow-Affected Watershed

摘要

Flood is one of the devastating natural disasters prediction of which is significantly important. Rainfall- runoffprocessandfloodingarephysicalphenomenathattheir investigation is very difficult due to effectiveness of differ- ent parameters. Various methods have been implemented to analyze these phenomena. The aim of current study is to investigate the performance of the artificial neural network (ANN) (hyperbolic tangent and sigmoid) and support vector machine (SVM) (regression type~(-1) and regression type-2) models to simulate the rainfall-runoff process influenced by snow water equivalent (SWE) height in Roodak water- shed, Tehran province, Iran. So, 92 MODIS images were gained from NASA website for three water years of 2003- 2005. Then, snow cover areas in all images were extracted and finally SWE values were calculated. Also, the data of precipitation, temperature and discharge for the mentioned yearswereusedformodeling.Accordingtotheresults,ANN with the hyperbolic tangent function, rainfall-temperature- SWE inputs, 1-day delay and RMSE and R~2 of 0.024 and 0.77, and the model with the sigmoid transfer function, rain- temperature-SWE inputs and RMSE and R~2 of 0.026 and 0.75 had better prediction capability than the other models. This indicates that the SWE has improved the accuracy of the models. The results of the SVM model indicate that the model with the rainfall-temperature-SWE, 1-delay, type-1 regression, RBF function and RMSE and R~2 of 0.054 and 0.030 had better prediction capability than other models. This also shows that consideration of the SWE enhances the performance and accuracy of the SVM models. More- over, comparing the results of ANN and SVM models, it can beconcludedthatANNmodelwiththerainfall-temperature- SWEinputs,1-daydelay,andthehyperbolictangentfunction had better predictions.