Wei et al. (2026) Evaluating and enhancing the performance of satellite precipitation products by considering uncertainty in rain gauge observations

Identification

• Journal: Natural Hazards
• Year: 2026
• Date: 2026-02-26
• Authors: Tai Wei, Xian-Ci Zhong, Yang Gao
• DOI: 10.1007/s11069-025-07936-3

Research Groups

• School of Civil Engineering and Architecture, Guangxi University
• Key Laboratory of Disaster Prevention and Structural Safety of Ministry of Education, Guangxi University
• Guangxi Engineering Research Center of Water Security and Intelligent Control for Karst Region, Guangxi University

Short Summary

This study develops a machine-learning-driven hierarchical framework to evaluate and correct biases in satellite precipitation products (SPPs) by incorporating uncertainty in rain gauge observations as interval-valued data. Applied to Guangxi, China, the framework significantly reduces SPP bias, especially for heavy precipitation events, through a novel evaluation index and a neural network-based correction model.

Objective

• To develop a machine-learning-driven hierarchical framework for evaluating and correcting biases in satellite precipitation products (SPPs), specifically by accounting for uncertainty in rain gauge observations using interval-valued data.

Study Configuration

• Spatial Scale: Guangxi, China (104°28′–112°04′ E longitude and 20°54′–26°23′ N latitude), characterized by complex topography including karst landforms. The study utilized data from 24 rain gauge stations and satellite products with a 0.1° spatial resolution.
• Temporal Scale: Satellite precipitation data (IMERG-E, IMERG-L, IMERG-F) from January 1, 2005, to December 31, 2014. Rain gauge observations from January 1, 2001, to December 31, 2015. All data were processed at a daily temporal scale. Bias correction focused on 2014 data.

Methodology and Data

• Models used:
  • Novel distance-based evaluation index ($R_I$) for interval-valued precipitation data.
  • Back-Propagation (BP) neural network for bias correction, featuring three hidden layers (20, 10, 10 neurons), Sigmoid activation function, a learning rate of 0.01, and the Levenberg–Marquardt (trainlm) algorithm.
• Data sources:
  • Satellite: Integrated Multi-satellitE Retrievals for GPM (IMERG) products (Early Run (IMERG-E), Late Run (IMERG-L), Final Run (IMERG-F)) Version 07, obtained from NASA’s official Early Earth Science center (http://disc.gshfc.nasa.gov/).
  • Observation: Daily precipitation records from 24 rain gauge stations in Guangxi, China, sourced from the China Meteorological Data Network (http://data.cma.cn).
  • Ancillary data for BP model: Elevation, longitude, latitude, daily maximum temperature, daily minimum temperature, and K-nearest neighbor satellite precipitation data.

Main Results

• A novel uncertainty-based performance index ($R_I$) was developed, classifying SPP performance into five hierarchies: Absolutely Accurate, Very Strongly Accurate, Strong Accuracy, Weak Accuracy, and Inaccuracy.
• Regional scale evaluation in Guangxi (2005-2014) showed that approximately 48% of IMERG data met acceptable standards ($R_I \ge 0.9$), while 52% required correction.
• SPP performance varied spatially, with higher accuracy in northwestern Guangxi and greater fluctuations in other areas, particularly coastal regions.
• For extreme precipitation events exceeding 50 mm, the original SPP performance was poor, with only 9.73% (IMERG-E), 8.79% (IMERG-L), or 10.98% (IMERG-F) falling into the acceptable accuracy range.
• The proposed BP neural network model significantly improved SPP accuracy. For torrential rain events (> 50 mm), the percentages of Absolutely Accurate, Very Strongly Accurate, and Strong Accuracy increased by 5.26%, 19.74%, and 15.79%, respectively, while Weak Accuracy and Inaccuracy decreased by 10.53% and 30.26%, respectively.
• Compared to traditional metrics, the corrected KBP-E data showed substantial improvements: Correlation Coefficient (CC) increased by 23.8%, Relative Bias (RB) decreased by 34.2%, Root Mean Square Error (RMSE) decreased by 30.7%, Probability of Detection (POD) increased by 43.9%, Critical Success Index (CSI) increased by 39.6%, and False Alarm Rate (FAR) decreased by 14.8%.
• The inclusion of K-nearest neighbor satellite precipitation data (specifically K=4) as input to the BP model enhanced correction capability, while excessive neighbors (K=8) led to redundancy and reduced accuracy.

Contributions

• Proposed the use of interval-valued observations to account for existing uncertainty in rain gauge measurements, with a detailed analysis of influencing factors specific to Guangxi, China.
• Developed a novel distance-based index for evaluating SPP performance, which was used to classify IMERG precipitation data and establish a hierarchical evaluation system.
• Established a bias correction rule based on the classified SPP performance and introduced a BP neural network model to approximate the complex functional relationship between satellite precipitation data and ground observations.
• Implemented a hierarchical training strategy for the BP neural network, which effectively corrected SPP biases, particularly for extreme precipitation events.
• Introduced a novel idea that the functional relationship between satellite precipitation data and rain gauge observations is consistent over time, departing from previous studies that often processed precipitation data as time series.

Funding

• National Natural Science Foundation of China (No. 11872155)

Citation

@article{Wei2026Evaluating,
  author = {Wei, Tai and Zhong, Xian-Ci and Gao, Yang},
  title = {Evaluating and enhancing the performance of satellite precipitation products by considering uncertainty in rain gauge observations},
  journal = {Natural Hazards},
  year = {2026},
  doi = {10.1007/s11069-025-07936-3},
  url = {https://doi.org/10.1007/s11069-025-07936-3}
}