Regionalization of hydrometeorological variables such as rainfall and temperature is necessary for various applications related to water resources planning and management. Sampling variability and randomness associated with the variables, as well as non-availability and paucity of data pose a challenge in modelling the variables. This challenge can be addressed by using stochastic models that utilize information from hydrometeorologically similar locations for modelling the variables. A set of locations that are hydrometeorologically similar are referred to as homogeneous region or pooling group and the process of identifying a homogeneous region is referred to as regionalization. The thesis concerns development of new approaches to regionalization of (i) extreme rainfall,(ii) maximum and minimum temperatures, and (iii) rainfall together with maximum and minimum temperatures. Regionalization of extreme rainfall and frequency analysis based on resulting regions yields quantile estimates that find use in design of water control (e.g., barrages, dams, levees) and conveyance structures (e.g., culverts, storm sewers, spillways) to mitigate damages that are likely due to floods triggered by extreme rainfall, and land-use planning and management. Regionalization based on both rainfall and temperature yield regions that could be used to address a wide spectrum of problems such as meteorological drought analysis, agricultural planning to cope with water shortages during droughts, downscaling of precipitation and temperature. Conventional approaches to regionalization of extreme rainfall are based extensively on statistics derived from extreme rainfall. Therefore delineated regions are susceptible to sampling variability and randomness associated with extreme rainfall records, which is undesirable. To address this, the idea of forming regions by considering attributes for regionalization as seasonality measure and site location indicators (which could be determined even for ungauged locations) is explored. For regionalization, Global Fuzzy c-means (GFCM) cluster analysis based methodology is developed in L-moment framework. The methodology is used to arrive at a set of 25 homogeneous extreme rainfall regions over India considering gridded rainfall records at daily scale, as there is dearth of regionalization studies on extreme rainfall in India Results are compared with those based on commonly used region of influence (ROI) approach that forms site-specific regions for quantile estimation, but lacks ability to delineate a geographical area into a reasonable number of homogeneous regions. Gridded data constitute spatially averaged rainfall that might originate from a different process (more synoptic) than point rainfall (more convective). Therefore to investigate utility of the developed GFCM methodology in arriving at meaningful regions when applied to point rainfall data, the methodology is applied to daily rainfall records available for 1032 gauges in Karnataka state of India. The application yielded 22 homogeneous extreme rainfall regions. Experiments carried out to examine utility of GFCM and ROI based regions in arriving at quantile estimates for ungauged sites in the study area reveal that performance of GFCM methodology is fairly close to that of ROI approach. Errors were marginally lower in the case of GFCM approach in analysis with observed point rainfall data over Karnataka, while its converse was noted in the case of analysis with gridded rainfall data over India. Neither of the approaches (CA, ROI) was found to be consistent in yielding least error in quantile estimates over all the sites. The existing approaches to regionalization of temperature are based on temperature time series or their related statistics, rather than attributes effecting temperature in the study area. Therefore independent validation of the delineated regions for homogeneity in temperature is not possible. Another drawback of the existing approaches is that they require adequate number of sites with contemporaneous temperature records for regionalization, because the delineated regions are susceptible to sampling variability and randomness associated with the temperature records that are often (i) short in length, (ii) limited over contemporaneous time period and (iii) spatially sparse. To address these issues, a two-stage clustering approach is developed to arrive at regions that are homogeneous in terms of both monthly maximum and minimum temperatures ( and ). First-stage of the approach involves (i) identifying a common set of possible predictors (LSAVs) influencing and over the entire study area, and (ii) using correlations of those predictors with and along with location indicators (latitude, longitude and altitude) as the basis to delineate sites in the study area into hard clusters through global k-means clustering algorithm. The second stage involves (i) identifying appropriate LSAVs corresponding to each of the first-stage clusters, which could be considered as potential predictors, and (ii) using the potential predictors along with location indicators (latitude, longitude and altitude) as the basis to partition each of the first-stage clusters into homogeneous temperature regions through global fuzzy c-means clustering algorithm. A set of 28 homogeneous temperature regions was delineated over India using the proposed approach. Those regions are shown to be effective when compared to an existing set of 6 temperature regions over India for which inter-site cross-correlations were found to be weak and negative for several months, which is undesirable. Effectiveness of the newly formed regions is demonstrated. Utility of the proposed maxTminT homogeneous temperature regions in arriving at PET estimates for ungauged locations within the study area was demonstrated. The estimates were found to be better when compared to those based on the existing regions. The existing approaches to regionalization of hydrometeorological variables are based on principal components (PCs)/ statistics/indices determined from time-series of those variables at monthly and seasonal scale. An issue with use of PCs for regionalization is that they have to be extracted from contemporaneous records of hydrometeorological variables. Therefore delineated regions may not be effective when the available records are limited over contemporaneous time period. A drawback associated with the use of statistics/indices is that they (i) may not be meaningful when data exhibit nonstationarity and (ii) do not encompass complete information in the original time series. Consequently the resulting regions may not be effective for the desired purpose. To address these issues, a new approach is proposed. It considers information extracted from wavelet transformations of the observed multivariate hydrometeorological time series as the basis for regionalization by global fuzzy c-means clustering procedure. The approach can account for dynamic variability in the time series and its nonstationarity (if any). Effectiveness of the proposed approach in forming homogeneous hydrometeorological regions is demonstrated by application to India, as there are no prior attempts to form such regions over the country. The investigations resulted in identification of 29 regions over India, which are found to be effective and meaningful. Drought Severity-Area-Frequency (SAF) curves are developed for each of the newly formed regions considering the drought index to be Standardized Precipitation Evapotranspiration Index (SPEI).

There is ample growth in scientific evidence about climate change. Since, hydrometeorological processes are sensitive to climate variability and changes, ascertaining the linkages and feedbacks between the climate and the hydrometeorological processes becomes critical for environmental quality, economic development, social well-being etc. As the river basin integrates some of the important systems like ecological and socio-economic systems, the knowledge of plausible implications of climate change on hydrometeorology of a river basin will not only increase the awareness of how the hydrological systems may change over the coming century, but also prepare us for adapting to the impacts of climate changes on water resources for sustainable management and development. In general, quantitative climate impact studies are based on several meteorological variables and possible future climate scenarios. Among the meteorological variables, sic “cardinal” variables are identified as the most commonly used in impact studies (IPCC, 2001). These are maximum and minimum temperatures, precipitation, solar radiation, relative humidity and wind speed. The climate scenarios refer to plausible future climates, which have been constructed for explicit use for investigating the potential consequences of anthropogenic climate alterations, in addition to the natural climate variability. Among the climate scenarios adapted in impact assessments, General circulation model(GCM) projections based on marker scenarios given in Intergovernmental Panel on Climate Change’s (IPCC’s) Special Report on Emissions Scenarios(SRES) have become the standard scenarios. The GCMs are run at coarse resolutions and therefore the output climate variables for the various scenarios of these models cannot be used directly for impact assessment on a local(river basin)scale. Hence in the past, several methodologies such as downscaling and disaggregation have been developed to transfer information of atmospheric variables from the GCM scale to that of surface meteorological variables at local scale. The most commonly used downscaling approaches are based on transfer functions to represent the statistical relationships between the large scale atmospheric variables(predictors) and the local surface variables(predictands). Recently Support vector machine (SVM) is proposed, and is theoretically proved to have advantages over other techniques in use such as transfer functions. The SVM implements the structural risk minimization principle, which guarantees the global optimum solution. Further, for SVMs, the learning algorithm automatically decides the model architecture. These advantages make SVM a plausible choice for use in downscaling hydrometeorological variables. The literature review on use of transfer function for downscaling revealed that though a diverse range of transfer functions has been adopted for downscaling, only a few studies have evaluated the sensitivity of such downscaling models. Further, no studies have so far been carried out in India for downscaling hydrometeorological variables to a river basin scale, nor there was any prior work aimed at downscaling CGCM3 simulations to these variables at river basin scale for various IPCC SRES emission scenarios. The research presented in the thesis is motivated to assess the impact of climate change on streamflow at river basin scale for the various IPCC SRES scenarios (A1B, A2, B1 and COMMIT), by integrating implications of climate change on all the six cardinal variables. The catchment of Malaprabha river (upstream of Malaprabha reservoir) in India is chosen as the study area to demonstrate the effectiveness of the developed models, as it is considered to be a climatically sensitive region, because though the river originates in a region having high rainfall it feeds arid and semi-arid regions downstream. The data of the National Centers for Environmental Prediction (NCEP), the third generation Canadian Global Climate Model (CGCM3) of the Canadian Center for Climate Modeling and Analysis (CCCma), observed hydrometeorological variables, Digital Elevation model (DEM), land use/land cover map, and soil map prepared based on PAN and LISS III merged, satellite images are considered for use in the developed models. The thesis is broadly divided into four parts. The first part comprises of general introduction, data, techniques and tools used. The second part describes the process of assessment of the implications of climate change on monthly values of each of the six cardinal variables in the study region using SVM downscaling models and k-nearest neighbor (k-NN) disaggregation technique. Further, the sensitivity of the SVM downscaling models to the choice of predictors, predictand, calibration period, season and location is evaluated. The third part describes the impact assessment of climate change on streamflow in the study region using the SWAT hydrologic model, and SVM downscaling models. The fourth part presents summary of the work presented in the thesis, conclusions draws, and the scope for future research. The development of SVM downscaling model begins with the selection of probable predictors (large scale atmospheric variables). For this purpose, the cross-correlations are computed between the probable predictor variables in NCEP and GCM data sets, and the probable predictor variables in NCEP data set and the predictand. A pool of potential predictors is then stratified (which is optional and variable dependant) based on season and or location by specifying threshold values for the computed cross-correlations. The data on potential predictors are first standardized for a baseline period to reduce systemic bias (if any) in the mean and variance of predictors in GCM data, relative to those of the same in NCEP reanalysis data. The standardized NCEP predictor variables are then processed using principal component analysis (PCA) to extract principal components (PCs) which are orthogonal and which preserve more than 98% of the variance originally present in them. A feature vector is formed for each month using the PCs. The feature vector forms the input to the SVM model, and the contemporaneous value of predictand is its output. Finally, the downscaling model is calibrated to capture the relationship between NCEP data on potential predictors (i.e feature vectors) and the predictand. Grid search procedure is used to find the optimum range for each of the parameters. Subsequently, the optimum values of parameters are obtained from the selected ranges, using the stochastic search technique of genetic algorithm. The SVM model is subsequently validated, and then used to obtain projections of predictand for simulations of CGCM3. Results show that precipitation, maximum and minimum temperature, relative humidity and cloud cover are projected to increase in future for A1B, A2 and B1 scenarios, whereas no trend is discerned with theCOMMIT. The projected increase in predictands is high for A2 scenario and is least for B1 scenario. The wind speed is not projected to change in future for the study region for all the aforementioned scenarios. The solar radiation is projected to decrease in future for A1B, A2 and B1 scenarios, whereas no trend is discerned with the COMMIT. To assess the monthly streamflow responses to climate change, two methodologies are considered in this study namely (i) downscaling and disaggregating the meteorological variables for use as inputs in SWAT and (ii) directly downscaling streamflow using SVM. SWAT is a physically based, distributed, continuous time hydrological model that operates on a daily time scale. The hydrometeorologic variables obtained using SVM downscaling models are disaggregated to daily scale by using k-nearest neighbor method developed in this study. The other inputs to SWAT are DEM, land use/land cover map, soil map, which are considered to be the same for the present and future scenarios. The SWAT model has projected an increase in future streamflows for A1B, A2 andB1 scenarios, whereas no trend is discerned with the COMMIT. The monthly projections of streamflow at river basin scale are also obtained using two SVM based downscaling models. The first SVM model (called one-stage SVM model) considered feature vectors prepared based on monthly values of large scale atmospheric variables as inputs, whereas the second SVM model (called two-stage SVM model) considered feature vectors prepared from the monthly projections of cardinal variables as inputs. The trend in streamflows projected using two-stage SVM model is found to be similar to that projected by SWAT for each of the scenarios considered. The streamflow is not projected to change for any of the scenarios considered with the one-stage SVM downscaling model. The relative performance of the SWAT and the two SVM downscaling models in simulating observed streamflows is evaluated. In general, all the three models are able to simulate the streamflows well. Nevertheless, the performance of SWAT model is better. Further, among the two SVM models, the performance of one-stage streamflow downscaling model is marginally better than that of the two-stage streamflow downscaling model.