A novel approach for constructing a grid-based precipitation dataset with uncertainty estimation using machine learning

Grid-based precipitation products play a crucial role in water resource management, particularly in data-scarce areas, leading to an upsurge of datasets being generated to fulfill this purpose. However, inherent inconsistency exists among these datasets due to variations in their approaches, tools, and data sources, resulting in considerable uncertainties in their applications. There has been very little attention given to the issue of quantifying confidence intervals in the current products. This study proposes a novel approach to this issue, producing a merged dataset with its confidence intervals over space and time attaining a spatial resolution of 0.1 degrees for the Vietnam land region over the period 2001-2010. Data gathered from 573 stations (covering 17% of the total grid cells) are merged with GPM IMERG and MERRA-2, supplemented by climatic variables such as wind and air temperature and additional auxiliary data (i.e., elevation, slope, and slope directions). Our estimated precipitation has improved precision in time and spatial distributions compared to its predecessors. We evaluate our product against up-to-date datasets specific to the Vietnam region, namely the Vietnam Gridded Precipitation. While the well-established merged Multisource Weighted-Ensemble Precipitation serves as the benchmarking method. We also demonstrate the value of the established confidence intervals (CIs) through statistical analyses which accentuate the remarkable dissimilarities in total precipitation and extreme events between the upper and lower bound of the CIs. These results further highlight the necessity of considering the confidence of constructed datasets in local applications. Ultimately, this study offers new perspectives for the future development of grid-based datasets, expanding beyond precipitation and into multiple dimensions.