Modeling Pan Evaporation by Hybrid Wavelet Transform Support Vector Machines

Abstract

Estimates of evaporation from open water bodies have gained a lot of importance and are crucial for several environment related functional systems, particular field related applications in water resources and ecology. Accurate estimations of evaporations would help these functional systems to overcome the crisis. The acute shortage of fresh water created serious issues in climatic zones like arid and semi-arid places. The creative and novel approaches are being employed in the estimation of evaporation. In this thesis, it is attempted to address the complex pan evaporation (PE) phenomenon by hybridizing Discrete Wavelet Transform (DWT) and Support Vector Machines (SVM). The superiority of DWT relies on its multi-resolution potential at various scales and SVM capable of establishing a rapid and accurate relationship between inputoutput patterns. Two stations, namely Bajpe (humid) and Bangalore (Semi- arid) located in the state of Karnataka, India are chosen for model development and to ensure efficiency of developed models. The model development begins SVM regression with kernel functions, namely Polynomial, Radial basis function (RBF) and Pearson VII function based kernel (PUK). The novel Gamma test (GT) was used to decide the best input output combination. Parameter optimization was carried by Grid search. The developed models showed better estimations of pan evaporation, but exhibited some limitations with non-linearity, sparse and noisy data. These limitations forced data pre-prcessing technique to get introduced via two mother functions, namely Daubechies (DB) and Haar.Fine tuning of various levels and orders were carried out. DB with order 3 and level 4 produced optimum results with SVR models. Overall, this hybrid combination shows promising potential to provide optimal solutions for problems arising out of pan evaporation estimation.