HR-PrecipNet: A machine learning framework for 1-km high-resolution satellite precipitation estimation

Abstract

Accurate and high-resolution precipitation estimation is critical for various applications in hydrology, meteorology, agriculture, and climate studies. This work proposes a novel machine learning (ML) framework for generating high-resolution (1 km) daily precipitation estimates over Italy by merging multi-source of information from top-down and bottom-up approaches. Our two-step framework firstly employs a deep learning (DL) architecture to produce initial 0.1-degree (approximately 10 km) daily precipitation estimates. We evaluate several U-Net DL architectures (2DCNN (Two-Dimensional Convolutional Neural Network), 3DCNN (Three-Dimensional CNN), ConvLSTM (Convolutional Long Short-Term Memory), Siamese, and Siamese-Diff), utilizing features such as infrared (IR), water vapor (WV) observation, soil moisture (SM), elevation, and geographical coordinates. Feature importance analysis underscores the significance of IR, WV, and differences in SM, demonstrating the value of integrating data from both approaches. The top-performing DL model achieves a correlation coefficient of 0.733 with ground-based data during the test period, a root mean square error of 4.06 mm, a bias close to zero, and a Critical Success Index (CSI) of 0.628. Secondly, we refine the estimates to 1 km resolution using a Random Forest (RF) model and high-resolution SM data. This refinement step crucially preserves the quality of the precipitation estimates. This approach effectively captures localized precipitation patterns across Italy and establishes a promising framework for the future development of 1-km high-resolution global precipitation products.