Author(s): S. Kannan; Subimal Ghosh
Keywords: Climate Change; Statistical Downscaling; Clustering; Classification And Regression Tree
Abstract: Hydrologic impacts of climate change at a river basin scale is normally assessed by simulating large scale climate variables with General Circulation Model (GCM) and downscaling the simulated climate variables to local scale climatic/hydrologic variables such as rainfall, evapotranspiration etc. . The downscaled variables are used in watershed models for future prediction of hydrologic scenario in a river basin. In statistical downscaling, statistical relationship is developed between the observed hydrologic variables and large scale climate variable and such relationship is further used with GCM simulated climate variables for prediction of hydrologic variables. For prediction of multi-site rainfall in a river basin, correlation between the rainfall of multiple sites is often not captured by conventional statistical downscaling technique. Furthermore these methods fail to model the variability of the predictand (hydrologic variable) and therefore going by the values of rainfall amount may not be always realistic. Therefore, it is proposed in the present study to represent the rainfall occurrence of all the station in a river basin, not by using values, but by using a rainfall state to develop the statistical downscaling model. The limitations of capturing correlations of stations with similar data can be overcome and at the same time, future rainfall condition can be estimated. An unsupervised clustering technique, K means clustering is used to derive the rainfall state from the multi-site rainfall data. Cluster validity measures are used to determine the optimum number of clusters. After obtaining the clusters a statistical relationship is obtained between the observed large scale climate data and the rainfall state (which are derived with K-means clustering). NCEP/NCAR reanalysis data have been used as a proxy to the observed large scale climate data. The relationship is obtained using classification and regression tree. The relationship thus developed is applied to the GCM simulated standardized bias free large scale climate variables for prediction of rainfall states in future. Comparisons of the number of days of different states for observed period and future will give the change expected in the river basin due to global warming. The methodology is demonstrated with Mahanadi river basin, India.