: Water scarcity, climate changes and hydrological uncertainties emphasize the necessity of a comprehensive and meaningful management of water resources which will be achieved by reliable models. Harnessing and appropriate using of surface water, developing the groundwater resources, alleviating the inverse impacts of flood or drought and healthy drinking water supply, require models with accurate prediction. Data-driven modeling is new and rapidly expanded in scientific and engineering research areas. This method, in some situations, would be a proper substitution for other modeling approaches such experimental and physical ones. In this research, the principles of a novel and advanced data-driven technique called Support Vector Machines (SVMs), which is relied on statistical learning theory, are discussed. By applying this learning method, generalization characteristic of the machine boosts; and thus, the model precision in comparison with other data-driven techniques improves. The aim of study is to present the underlying concepts of SVMs in order to learn complex physical processes and non-linear behaviors of hydrological systems. In this thesis, the performance of learning machine is investigated within three various applications of water resources modeling consisting of 1) prediction of short-term runoff, 2) estimating the water level of a specific observation well, and 3) spatial-temporal forecasting of rainfall. For implementing SVMs in the foregoing hydrological systems, diverse combination of data is employed. Analyzing and comparing the SVM results with the ones obtained through the artificial neural networks reveals the high prediction capability of SVM in the preceding applications. The successful performance of SVMs throughout this research signifies the possibility of exploiting it in other water resources applications.
