Improving near-real-time satellite precipitation products through multistage modified schemes

Near-real-time (NRT) satellite precipitation products (SPPs) exhibit great application potential in global precipitation estimation for their convenient acquisition and access. A universal framework was developed to modify precipitation occurrence and intensity in stages to correct and integrate multiple NRT SPPs with open access gridded observational data. Based on this framework, both machine learning-and statistical-based correction methods were utilized to construct multi-stage modified schemes and three schemes, namely the double machine learning (DML), statistical error correction (SEC) and hybrid (ML-SEC) schemes, were stepwise established. The improvements of NRT SPPs by these three schemes were comparatively analyzed based on a dense network of rain gauges located in the Yiluo River Basin, China, which does not depend on reference data for modification and has derived a reliable evaluation result at grid-point scale. The results indicated that the DML scheme attains the highest accuracy with a grid-averaged Kling-Gupta efficiency (KGE) value of 0.70 and a critical success index (CSI) value of 0.60, followed by the ML-SEC and SEC schemes. The KGE and CSI values of NRT SPPs have been improved by 22.20% 535.48% and 30.30% 100.24%, respectively, through the DML scheme. The DML-modified result also outperformed the mainstream post-processed SPPs with a 1.95% 222.25% and 9.22% 102.00% higher KGE and CSI values, respectively. The main error sources of the SPPs due to false and hit events were reduced by the multistage schemes, and the DML scheme achieved a balanced efficiency at both occurrence and intensity modification stages. Moreover, the DML scheme was more advantageous than spatial interpolation under the scenario with fewer gauges, particularly in estimating precipitation occurrence. The multistage modified schemes in this study, especially the DML scheme, exhibit potential in near real time and sub-daily scale applications. Multistage schemes for NRT SPP modification are helpful for obtaining high-quality precipitation estimates.