Monday, 29 January 2024
Hall E (The Baltimore Convention Center)
Accurate quantification of precipitation (QPE) is of utmost importance for a wide range of applications, including flood forecasting, drought monitoring, and land surface modeling. However, there exists diverse uncertainty among in situ precipitation datasets, remote sensing-based estimations, and reanalysis products. Many models have been developed to merge these estimations from different sources, aiming to enhance QPE accuracy. However, most of these attempts focus primarily on spatial or temporal correlations between remote sensing and gauge data separately, which limits their ability to fully capture the underlying spatiotemporal dependencies that could lead to better precipitation estimations. In this study, we have introduced a comprehensive framework capable of simultaneously merging and downscaling multiple user-defined precipitation products using rain gauge observations as reference values. To achieve this, we designed an innovative deep learning-based convolutional neural network architecture, referred to as the precipitation data fusion network (PDFN). The PDFN incorporates several layers of 3D-CNN and ConvLSTM, allowing us to exploit the spatial and temporal patterns of precipitation effectively. The framework, built using Amazon Web Services (AWS), automates pre-processing of remote sensing and in-situ data, execution of computationally expensive deep learning models, and visualization of the final precipitation product. The system is designed to run automatically to produce high-resolution precipitation estimation over CONUS on a daily basis.

