A Machine Learning Approach for Improving the Accuracy of Gridded Precipitation With Uncertainty Quantification

Abstract

The proliferation of gridded precipitation datasets produced through diverse methods has led to user confusion due to discrepancies in values for identical locations and times, underscoring the inherent uncertainty in current precipitation data. However, quantifying this uncertainty remains challenging since most datasets are deterministic and offer no easy mechanism for such quantification. This study proposes a novel machine learning (ML) approach that involves merging multiple satellite‐based gridded precipitation datasets and rain gauge observations. The ultimate goal is to create an improved high‐resolution (0.1°) daily precipitation product for Vietnam (2001–2010) that includes uncertainty quantification. By combining eXtreme gradient boosting (XGB) with quantile regression, we generate both deterministic estimates along with the associated uncertainty intervals. The resulting dataset, VNpu (Vietnam Precipitation with Uncertainty), outperforms individual input products, benchmark interpolation methods and an existing gauge‐based product (VnGP), particularly for heavy and extreme rainfall events. The VNpu provides spatiotemporally varying uncertainty ranges, with larger uncertainties in sparse gauge coverage areas. The analysis of ML interpretability reveals the complementary nature of multiple precipitation inputs (IMERG and MERRA2) and auxiliary topographic information. This study not only highlights the necessity of gridded precipitation products with uncertainty estimates but also demonstrates the value of ML approaches in developing improved precipitation products for data‐scarce regions.

Bibliographic Information