An Appraisal of Dynamic Bayesian Model Averaging-based Merged Multi-Satellite Precipitation Datasets Over Complex Topography and the Diverse Climate of Pakistan

Abstract

Merging satellite precipitation products tends to reduce the errors associated with individual satellite precipitation products and has higher potential for hydrological applications. The current study evaluates the performance of merged multi-satellite precipitation dataset (daily temporal and 0.25° spatial resolution) developed using the Dynamic Bayesian Model Averaging algorithm across four different climate regions, i.e., glacial, humid, arid and hyper-arid regions, of Pakistan during 2000–2015. Four extensively evaluated SPPs over Pakistan, i.e., Tropical Rainfall Measurement Mission (TRMM) Multi-satellite Precipitation Analysis (TMPA) 3B42V7, Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR), Climate Prediction Center MORPHing technique (CMORPH), and Era-Interim, are used to develop the merged multi-satellite precipitation dataset. Six statistical indices, including Mean Bias Error, Mean Absolute Error, Root Mean Square Error, Correlation Coefficient, Kling-Gupta efficiency, and Theil’s U coefficient, are used to evaluate the performance of merged multi-satellite precipitation dataset over 102 ground precipitation gauges both spatially and temporally. Moreover, the ensemble spread score and standard deviation are also used to depict the spread and variation of precipitation of merged multi-satellite precipitation dataset. Skill scores for all statistical indices are also included in the analyses, which shows improvement of merged multi-satellite precipitation dataset against Simple Model Averaging. The results revealed that DBMA-MSPD assigned higher weights to TMPA (0.32) and PERSIANN-CDR (0.27). TMPA presented higher skills in glacial and humid regions with average weights of 0.32 and 0.37 as compared to PERSIANN-CDR of 0.27 and 0.25, respectively. TMPA and Era-Interim depicted higher skills during pre-monsoon and monsoon seasons, with average weights of 0.31 and 0.52 (TMPA) and 0.25 and 0.21 (Era-Interim), respectively. Merged multi-satellite precipitation dataset overestimated precipitation in glacial/humid regions and showed poor performance, with the poorest values of mean absolute error (2.69 mm/day), root mean square error (11.96 mm/day), correlation coefficient (0.41), Kling-Gupta efficiency score (0.33) and Theil’s U (0.70) at some stations in glacial/humid regions. Higher performance is observed in hyper-arid region, with the best values of 0.71 mm/day, 1.72 mm/day, 0.84, 0.93, and 0.37 for mean absolute error, root mean square error, correlation coefficient, Kling-Gupta Efficiency score, and Theil’s U, respectively. Merged multi-Satellite Precipitation Dataset demonstrated significant improvements as compared to TMPA across all climate regions with average improvements of 45.26% (mean bias error), 30.99% (mean absolute error), 30.1% (root mean square error), 11.34% (correlation coefficient), 9.53% (Kling-Gupta efficiency score) and 8.86% (Theil’s U). The ensemble spread and variation of DBMA-MSPD calculated using ensemble spread score and standard deviation demonstrates high spread (11.38 mm/day) and variation (12.58 mm/day) during monsoon season in the humid and glacial regions, respectively. Moreover, the improvements of DBMA-MSPD quantified against fixed weight SMA-MSPD reveals supremacy of DBMA-MSPD, higher improvements (40–50%) in glacial and humid regions.