Bias correction of global circulation model outputs using artificial neural networks

Bias correction;Regression models;Artificial neural networks;Generalization;Regionalization;Community climate system model;Climate data sets

Moghim, Sanaz ; Bras, Rafael L. Georgakakos, Aris P. Wang, Jingfeng Hsu, Kuo-lin Deng, Yi Civil and Environmental Engineering ; Bras, Rafael L.

University:Georgia Institute of Technology

Department:Civil and Environmental Engineering

关键词: Bias correction;    Regression models;    Artificial neural networks;    Generalization;    Regionalization;    Community climate system model;    Climate data sets;   

Others  :  https://smartech.gatech.edu/bitstream/1853/55487/1/MOGHIM-DISSERTATION-2015.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

Climate studies and effective environmental management plans require unbiased climate datasets. This study develops a new bias correction approach using a three layer feedforward neural network to reduce the biases of climate variables (temperature and precipitation) over northern South America. Air and skin temperature, specific humidity, net longwave and shortwave radiation are used as inputs for the bias correction of temperature. Precipitation at lag zero, one, two, and three, and the standard deviation from 3 by 3 neighbors around the pixel of interest are the inputs into the ANN bias correction of precipitation. The data are provided by the Community Climate System Model (CCSM3). Results show that the trained ANN can markedly reduce the estimation error and improve the correlation and probabilistic structure of the bias-corrected variables for calibration and validation periods. The ANN outperforms linear regression (LR), which is used for comparison purposes. The ability of the regression models (linear and ANN) to regionalize the study domain is investigated by defining the minimum number of training pixels necessary to achieve a good level of bias correction performance over the entire domain. Results confirm that it is possible to identify regions in terms of physical features such as land cover, topography, and climatology over which the trained models at a few pixels can do well. The new approach saves computational demands, time, and memory usage and it can be used for other climate models efficiently.