Representative grid location-multivariate adaptive regression spline (RGL-MARS) algorithm for downscaling dry and wet season rainfall

• Published in: Journal of hydrology (Amsterdam) Vol. 605; p. 127381
• Main Authors: Panda, Kanhu Charan, Singh, R.M., Thakural, L.N., Sahoo, Debi Prasad
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-02-2022
• Subjects: Decision tree; General circulation models; Multiple linear regression; Multivariate adaptive regression spline; Representative grid location; Statistical downscaling; Statistical downscaling; Representative grid location; Multiple linear regression; General circulation models; Multivariate adaptive regression spline; Decision tree
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2021.127381

•Advocated a novel downscaling model to capture the nonlinearity in the rainfall.•Designed a framework to evaluate the developed model with existing models.•Developed model outperformed the existing models during both dry and wet seasons.•Extreme events are captured efficiently employing the developed model.
The high dimensionality of predictor variables reduces the predictive accuracy of statistical downscaling models. Principal component analysis (PCA) is one of the extensively used approaches for reducing the dimensionality of the predictors. However, PCA reduces the efficiency of downscaling models when a nonlinear predictor-predictand relationship exists. To solve this issue, the representative grid location (RGL) approach was used to minimise the dimension of the predictor variables. Thus, a novel RGL-MARS (Multivariate Adaptive Regression Spline) based downscaling model was proposed in the current study. The proposed model was compared with PCA-MLR (Principal Component Analysis-Multiple Linear Regression), MARS, and PCA-MARS downscaling methods. Eight general circulation models (GCMs) were considered, out of which only CAN-ESM2 (second-generation Canadian Earth System Model) GCM was found suitable for the study area. Three criteria, i.e., correlation coefficient (CC), mutual information (MI), and decision tree (DT), were used for the selection of dominant predictors. It was observed that CMD (CC + MI + DT) predictors, when used with the RGL-MARS downscaling method, showed the best performance (NSE = 74–89%; root mean square error (RMSE) = 10–40 mm). The results revealed that rainfall simulated by the RGL-MARS model captured the standard deviation and coefficient of variation in the observed rainfall, whereas the rest failed to do so. The RGL-MARS model solved the issue of underprediction of wet season rainfall and overprediction of dry season rainfall. The RGL-MARS addressed the issue with an improved NSE of 62–89% and RMSE of 18–45 mm; thus, the result exhibited that integration of MARS with the RGL perform better compared to the PCA.. This study also demonstrated that downscaling model outcomes were more reliable for the wet than the dry season. The future rainfall projection indicated that the rainfall fluctuation might be more in the dry season than the wet season. The proposed downscaling method can improve the accuracy of rainfall projections under various climatic conditions, subsequently, coupled with a range of hydrological and land surface models to better understand the catchment characteristics and the water balance dynamics in future climate studies.