STUDY ON STATISTICAL DOWNSCALING FOR CLIMATE- PROJECTIONS

Abstract

General Circulation Models (GCMs) are the most powerful tools available to simulate evolving and future changes in the climate system. GCMs are numerical models that represent the large-scale physical processes of the earth-atmosphere-ocean system and have been designed to simulate the past, present, and future climate. 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 (lake/river basin) scale. The spatial scale on which a GCM can operate (e.g., 2.5° longitude x 2.5 ° latitude for Global Climate Model, CGCM3) is very coarse compared to that of a hydrologic process (e.g., precipitation in a region etc.) of interest in the climate change impact assessment studies. Moreover, accuracy of GCMs, in general, decreases from climate related variables, such as temperature and air pressure to hydrologic variables such as precipitation, evapotranspiration, which are also simulated by GCMs. These limitations of the GCMs restrict the direct use of their output in hydrology. Hydrologic variables, such as temperature and precipitation, etc., are significant parameters for climate change impact studies. A proper assessment of probable future temperatures and precipitation and their variability should be done for various hydro-climatology scenarios. The methods used to convert GCM outputs into local meteorological variables required for reliable hydrological modeling are usually referred to as `downscaling' techniques. These techniques 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). The literature review revealed that there have been several downscaling approaches proposed in the several studies, each having its own advantages and shortcomings. It is not obvious which downscaling technique provides the most reliable simulation of the hydro-climatic variables since there are factors such as the topography of a region which can influence the performance of a downscaling model. Therefore, a rigorous evaluation of the various downscaling techniques must be performed to identify the most robust method(s) for a given location. This comprehensive study explores the applicability of using various statistical and machine learning approaches for downscaling hydrometeorological variables to a lake/river basin scale. The statistical methods investigated in this study are multiple linear regressions such as direct, forward, backward, stepwise, locally weighted regression ii smoothing scatter plots technique (LOWESS), robust version of LOWESS, K-nearest neighbor and partial least squares (PLS) regression while Artificial Neural Networks (ANNs) and decision tree algorithms such as Single Conjunctive Rule Learner, Decision Table, M5 Model Tree, and REPTree are explored as a downscaling tool. The selection of appropriate predictors is usually done using cross correlation and scatter plot methodology. Variable selection with the variable importance in the projection (VIP) scores provide a measure of importance of each explanatory variable or predictor and as a measure of performance when multi collinearity exists among variables through computer simulation experiments. No studies so far have used VIP score to select the predictors in downscaling exercise. A major limitation of K-nearest neighbor based weather generators is that they do not produce new values but merely reshuffle the historical data to generate realistic weather sequences and also generates negative precipitation amounts using normal kernels which is highly undesirable. Thus, there is a need of approaches which do not produce negative precipitation. Downscaled variables using GCM are often characterized by biases that limit their direct application for basin level hydrological modeling and several bias correction approaches are in use. In this thesis, two study areas namely Pichola lake basin and Upper Thames River Basin (UTRB) have been used which are located in the and region in India and humid region in Canada, respectively. The data used for this study consists of daily and monthly atmospheric variables simulated by Canadian Center for Climate Modeling and Analysis's (CCCma), third generation Coupled Global Climate Model (CGCM3). The data comprises of 20th century simulations (20C3M), and future simulations forced by four SRES scenarios namely, A1B, A2, B1 and COMMIT for the period 2001-2100. Reanalyzed data of the daily/monthly mean atmospheric variables prepared by NCEP/NCAR are used. In order to relate the large-scale weather patterns to the local scale, downscaling is necessary. The data of potential predictors is first standardized. Standardization is widely used prior to statistical downscaling to reduce bias (if any) in the mean and the variance of GCM predictors with respect to that of NCEP-reanalysis data. To develop the downscaling model, the feature vectors (i.e. predictors) which are prepared from NCEP record are iii partitioned into a training set and a test set. The different statistical parameters of each model are calculated during calibration to get the best statistical agreement between observed and simulated meteorological variables. For this purpose, various statistical performance measures, such as Coefficient of Correlation (CC), Standard Error of Estimate (SSE), Mean Square Error (MSE), Root Mean Square Error (RMSE), Normalized Mean square Error (NMSE), Nash—Sutcliffe Efficiency Index and Mean Absolute Error (MAE) were used to measure the performance of various models. Once the downscaling models have been calibrated and validated, the next step is to use these models to downscale the scenarios simulated by the GCM. The research reported in this thesis contributes towards applying several methodologies for projecting hydrologic variables such as temperature, precipitation at local scale from large scale GCM output in Pichola lake basin, Rajasthan, India in and region and Upper Thames River Basin (UTRB), Ontario, Canada in the humid region. All methodologies described earlier have been applied extensively for Pichola lake basin in India and identified best methodologies have been used for UTRB in Canada. The large scale atmospheric variables simulated by the third generation Canadian GCM for various IPCC scenarios (SRES A1B, SRES A2, SRES B 1 and COMMIT) were used to prepare inputs to models for downscaling purposes. GCM bias correction procedure improved the overall predictability of predictands. For selection of predictors, methodologies namely (i) cross correlation and scatter plot (ii) variable importance in projection (VIP) score were used. The novel approach of selection of predictors using VIP score yielded similar results that of using cross correlation and scatter plot. Several methodologies have been applied for downscaling Tmax, Tmin, precipitation and evaporation on monthly values for Pichola lake basin using simulations of CGCM3. Overall, model developed using M5 model tree and artificial neural network (ANN) performed better than any other model using other approaches investigated here in both calibration as well as validation for all predictands (precipitation, Tmax, Tmin and evaporation). The models developed using RLOWESS approach performed worst among the all other models developed using other approaches for predictand Tmax and Tmin as well as precipitation while PLS regression performed worst for predictand evaporation. The results of downscaling show that Tmax and Tmin are projected to increase in future for A1B, A2 and B1 scenarios, whereas no trend is discerned with the COMMIT. The projected increase in predictands is high for A2 scenario, whereas it is least for B 1 scenario. For pan iv evaporation, it can be concluded that trend is not obvious for future years since the factors working on pan evaporation are complicated. The results of downscaling models show that precipitation is projected to increase in future for A2 scenario, whereas there is no obvious trend for other scenarios. For the Upper Thames River Basin (UTRB), the suitability of various available rule and decision tree learning algorithms and artificial neural network (ANN) approaches were explored to downscale Tmax and Tmin as well as precipitation from CGCM3 outputs to local scale across 14 stations on daily scale. Overall, model developed using M5 model tree performed better than ANN model, investigated here in both calibration as well as validation for all predictands (Tmax and Tmin as well as precipitation). It has been observed that there is an increasing trend for the period 2081-2100 and all the scenarios considered (i.e. A1B, A2 and B 1) for all stations for predictands Tmax and Tmin and there is an increasing trend for the period 2046-2065 for all the scenarios for predictand precipitation. A K-nearest neighbour algorithm based on gamma kernel perturbation was developed which enables generation of data that are not the same as the historical data and gamma kernel may be a better choice for hydrological data being bounded by zero. An important aspect of the proposed model is that extreme events, such as high precipitation, can be simulated. This may be valuable aid in flood prediction models if their performance is evaluated based on synthetic sequences generated by the proposed model. Additionally, no site specific assumptions regarding the probability distribution of variables are required. v