Decision tree-based reduction of bias in monthly IMERG satellite precipitation dataset over India

Abstract

Decision trees are ideally suited for handling huge datasets and modelling non-linear relationships between different variables. Given the relationship between precipitation and bias may be very complex and non-linear, bias-correction of satellite precipitation is a challenge. We examine the applicability of Classification and Regression tree (CART) for bias-correction of the Integrated Multi-satellite Retrievals for Global Precipitation Mission (IMERG) precipitation dataset over India. The gauge-based 0.25° gridded precipitation dataset from India Meteorological Department is considered as the reference. The CART model is trained (2001–2011) and tested (2012–2016) over each 0.25° grids. The training dataset is subjected to 10-fold cross-validation and optimization of the minimum size of leaf node (one of the hyper-parameter). Efficiency of the CART model is evaluated using performance metrics like R2, RMSE and MAB over the whole of India and different climate and elevation zones in India. CART model is observed to be highly effective in capturing the bias during the training (average R2 ¼ 0.77) and testing (average R2 ¼ 0.66) period. Significant improvement in average monthly MAB (-6.3 to 29.2%) and RMSE (8.7–37.3%) was obtained post bias-correction by CART. Better performance of CART model was observed when compared to two widely adopted bias-correction techniques.