Accurate precipitation data are crucial in atmospheric and hydrological studies, especially for water resource management and disaster early warning. Satellite precipitation product (SPP) with high spatiotemporal resolution has been regarded as a valuable alternative precipitation source to ground observations. However, the hourly SPP generally performs poorly compared to daily SPP, thereby bias correction is urgently required. This study investigates the viability of utilizing machine learning methods to correct the bias of the hourly Integrated Multi-satellitE Retrievals for Global Precipitation Measurement-Early (IMERG-E) product. Meanwhile, the Weather Research and Forecasting (WRF) model is utilized to generate high-resolution fields of four hourly meteorological variables, namely, temperature at 2 m (TEMP2), specific humidity at 2 m (Q2), wind direction at 10 m (WDIR10), and wind speed at 10 m (WSPD10), which further serve as covariates in machine learning models to enhance the correction process. Four machine learning models were developed, i.e., Random Forest (RF) and Bidirectional Long Short-Term Memory Networks (Bi-LSTM) without WRF-simulated covariates, and RF-WRF and Bi-LSTM-WRF with meteorological covariates. The results demonstrated that incorporating WRF-simulated meteorological covariates improved model performance. Specifically, correlation coefficient (CC) values increased from 0.47 (RF) to 0.51 (RF-WRF) and rose from 0.55 (Bi-LSTM) to 0.60 (Bi-LSTM-WRF), along with reduced root mean square error (RMSE) and increased critical success index (CSI) values. Furthermore, two Bi-LSTM models consistently outperformed two RF models. Overall, the Bi-LSTM-WRF model emerged as the most effective correction method, which increased CC from 0.43 (IMERG-E) to 0.60, reduced RMSE from 1.91 mm to 1.08 mm, and enhanced CSI from 0.34 to 0.41. This study underscores the potential of integrating high-resolution WRF meteorological outputs into machine learning frameworks for correcting hourly SPPs, contributing significantly to the advancement of precipitation estimation in meteorological and hydrological applications.

中文翻译：

---

使用高分辨率 WRF 气象模拟增强的机器学习方法对每小时卫星降水产品进行偏差校正


准确的降水数据对于大气和水文研究至关重要，特别是对于水资源管理和灾害预警。具有高时空分辨率的卫星降水产品（SPP）被认为是地面观测的有价值的替代降水源。然而，与每日SPP相比，每小时SPP通常表现较差，因此迫切需要进行偏差修正。本研究调查了利用机器学习方法纠正每小时全球降水测量早期综合多卫星检索 (IMERG-E) 产品偏差的可行性。同时，利用天气研究与预报（WRF）模型生成四个每小时气象变量的高分辨率场，即2 m处的温度（TEMP2）、2 m处的比湿度（Q2）、10 m处的风向（ WDIR10) 和 10 m 风速 (WSPD10)，进一步作为机器学习模型中的协变量来增强校正过程。开发了四种机器学习模型，即没有 WRF 模拟协变量的随机森林 (RF) 和双向长短期记忆网络 (Bi-LSTM)，以及具有气象协变量的 RF-WRF 和 Bi-LSTM-WRF。结果表明，纳入 WRF 模拟的气象协变量可提高模型性能。具体而言，相关系数 (CC) 值从 0.47 (RF) 增加到 0.51 (RF-WRF)，从 0.55 (Bi-LSTM) 增加到 0.60 (Bi-LSTM-WRF)，同时均方根误差 (RMSE) 降低并增加关键成功指数（CSI）值。此外，两个 Bi-LSTM 模型的性能始终优于两个 RF 模型。总体而言，Bi-LSTM-WRF 模型成为最有效的校正方法，它将 CC 从 0.43 (IMERG-E) 提高到 0。60，将 RMSE 从 1.91 mm 降低至 1.08 mm，并将 CSI 从 0.34 提高至 0.41。这项研究强调了将高分辨率 WRF 气象输出集成到机器学习框架中以校正每小时 SPP 的潜力，从而为气象和水文应用中降水估算的进步做出重大贡献。