Quantile-based Bayesian Model Averaging approach towards merging of
rainfall products
Abstract
Due to the advancement in satellite and remote sensing technologies, a
number of satellite precipitation products (SPPs) are easily accessible
online at free of cost. These precipitation products have a huge
potential for hydro-meteorological applications in data-scare
catchments. However, the use of such products is still limited owing to
their lack of accuracy in capturing the ground truth. To improve the
accuracy of these products, we have developed a quantile based Bayesian
model averaging (QBMA) approach to merge the satellite precipitation
products. QBMA is a probabilistic approach to assign optimal weights to
the SPPs depending on their relative performances. The QBMA approach is
compared with simple model averaging and one outlier removed. TRMM,
PERSIANN-CDR, CMORPH products were experimented for QBMA merging during
the monsoon season over India’s coastal Vamsadhara river basin. QBMA
optimal weights were trained using 2001 to 2013 daily monsoon rainfall
data and validated for 2014 to 2018. Results indicated that QBMA
approach with bias corrected precipitation inputs outperformed the other
merging methods. On monthly evaluation, it is observed that all the
products perform better during July and September than that in June and
August. The QBMA approaches do not have any significant improvement over
the SMA approach in terms of POD. However, the bias-corrected QBMA
products have lower FAR. The developed QBMA approach with bias-corrected
inputs outperforms the IMERG product in terms of RMSE.