Correcting bias of satellite rainfall data using physical empirical model

The provision of high resolution near real-time rainfall data has made satellite rainfall products very potential for monitoring hydrological hazards. However, a major challenge in their direct-use can be problematic due to measurement error. In this study, an attempt was made to correct the bias of Global Satellite Mapping of Precipitation near-real-time (GSMaP_NRT) product. Physical factors, including topography, season, windspeed and cloud types were accounted for correcting bias. Peninsular Malaysia was used as the case study area. Gridded rainfall, developed from 80 gauges for the period 2000-2018, was used along with physical factors in a two-stage procedure. The model consisted of a classifier to categorise rainfall of different intensity and regression models to predict intensity class of different rainfall amount. An ensemble tree-based learning algorithm, called random forest, was used for classification and regression. The results revealed a big improvement of near-real-time GSMaP_NRT product after bias correction (GSMaP_BC) compared to the gauge corrected version (GSMaP_GC). Accuracy evaluation for complete time series indicated about 110% reduction of normalized root-mean-square error (NRMSE) in GSMaP_BC (0.8) compared to GSMaP_NRT (1.7) and GSMaP_GC (1.75). On the other hand, the bias of GSMaP_BC became nearly zero (0.3) compared to 2.1 and -3.1 for GSMaP_NRT and GSMaP_GC products. The spatial correlation of GSMaP_BC was >0.7 with observed rainfall data for all months compared to 0.20.78 for GSMaP_NRT and GSMaP_GC, indicating capability of GSMaP_BC to replicate spatial pattern of rainfall. The bias-corrected near-real-time GSMaP data can be used for monitoring and forecasting floods and hydrological phenomena in the absence of dense rain-gauge network in areas, frequently experience hydro-meteorological hazards.