Md. Bazlur Rashid*
Bangladesh Meteorological Department, Agargaon, Dhaka-1207, Bangladesh
Syed Shahadat Hossain
Institute of Statistical Research and Training (ISRT), University of Dhaka, Dhaka-1000, Bangladesh
Climate Change; Post-monsoon Season; Empirical Statistical Downscaling; GCM.
Resource Development and Management
Modeling, Temperature
2.1 Data description The analysis in this paper is based on mean temperature data from 15 GCM simulations from the fifth phase of the Coupled Model Inter-comparison Project (CMIP5) ensemble. The GCM data were downloaded from the KNMI Climate Explorer (https://climexp.knmi.nl/start.cgi) which offers data re-gridded to a 2.5- degree resolution grid for the time period of 1900-2100. Three future scenarios (Representative Concentration Pathways, RCPs) were considered: the high emission scenario RCP8.5 (Riahi et al., 2007), the medium emission scenario RCP4.5 in which the radiative forcing stabilizes shortly after 2100 (Clarke et al., 2007), and the more optimistic peak-and-decline scenario RCP2.6 (Van Vuuren et al., 2007). Local observations of the mean temperature from stations in Bangladesh were observed local variables data and are collected from the Bangladesh Meteorological Department (BMD). The ERA- Interim (resolution 79 km) reanalysis data set is a global atmospheric reanalysis produced by the European Centre for Medium?Range Weather Forecasts (ECMWF). The ERA-Interim project was produced in part to prepare for a new atmospheric reanalysis to replace ERA-40.
2.2 Data Analysis Statistical downscaling first generates a statistical relationship between larger GCM scale variables and observed small-scale (station level) variables. Different approaches can be used such as analog methods (rotation typing), regression analysis, or neural network techniques (Wilby et al., 2002). Future values of the large-scale variables found from GCM projections of future climate are then used to drive the statistical relationships in order to estimate the smaller-scale particulars of future climate. Statistical models usually consist of equations as shown below-
The empirical-statistical downscaling approach used in this study incorporated a form for quality control and bias adjustment through the practice of common Empirical Orthogonal Functions (EOFs) in the representation of the large-scale predictors (Benestad et al., 2015; 2016). After bias correction, a stepwise multiple regression was used to estimate model parameters, hence downscaling large-scale climate variables to local scale. Such a statistical downscaling approach requires a smaller amount of computational effort than dynamic downscaling and can be applied to many scenarios and longtime intervals, rather than the short-term slices of the dynamical downscaling method.
In this study Principal Component Analysis (PCA) was applied to the observational data before downscaling. The PCA decomposed the data into a set of spatial patterns, corresponding time series that describe the temporal variability associated with each pattern, and eigenvalues that represent the relative strength of each pattern. Rather than downscaling the observational stations individually, the time series associated with the first spatial patterns (hereby referred to as first principal components) were downscaled. The projected temperature could then be reconstructed from the downscaled principle components combined with the corresponding spatial patterns and eigenvalues.
2.3 Analysis Software The study was carried out within the R-environment and used Empirical Statistical Downscaling (ESD) package (Benestad et al., 2015) to analyze the data for attaining the objectives. The development of the ESD software fits in with the trend of the R-language increasing role in the climate change debate and as an open science platform. Additionally, both R and the ESD R-package are appreciated tools for linking high education and research. The wide range of functionalities of the ESD tool, including methods for reading and manipulating data, generating various info-graphics, and performing statistical analysis (e.g., calculating EOF), PCA, canonical correlation analysis (CCA), and empirical-statistical downscaling create it appropriate for processing results from GCMs.
Journal of Engineering Science 11(2), 2020, 27-35
Journal