Bias correction of satellite and reanalysis products for daily rainfall occurrence and intensity

2025-10-31

AI Summary

To address systematic biases in satellite and reanalysis precipitation estimates (SREs) over data-scarce regions, this study systematically evaluates bias correction methods—including statistical approaches (LOCI, quantile mapping), machine learning models (SVR, GPR), and hybrid techniques (LOCI-GPR, QM-GPR)—for daily-scale correction of seven SREs. Validation across >70 ground stations in Ghana and Zambia reveals that ENACTS, which assimilates dense in situ observations, exhibits the highest correctability. Overall, statistical methods outperform standalone machine learning models. Bias-corrected products achieve dry-day detection probabilities ≥0.80, substantially improving drought characterization; however, detection probability for extreme rainfall events remains ≤0.20, exposing a shared limitation in modeling the distributional tails. This work is the first to identify an intrinsic asymmetry in SRE bias correction—namely, “easier drought identification but harder extreme-precipitation identification”—providing critical guidance for method selection and algorithmic improvement in high-stakes hydro-meteorological applications.

Technology Category

Machine Learning: Calibration & Uncertainty QuantificationIntelligent Robots: State EstimationData Mining & Knowledge Management: Anomaly/Outlier Detection

Application Category

Search and Retrieval-Augmented AI: Web evaluation methodologies and metricsSocial Networks and Social Media: Fairness and bias in social network and social media analysisSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Satellite and reanalysis rainfall products (SREs) can serve as valuable complements or alternatives in data-sparse regions, but their significant biases necessitate correction. This study rigorously evaluates a suite of bias correction (BC) methods, including statistical approaches (LOCI, QM), machine learning (SVR, GPR), and hybrid techniques (LOCI-GPR, QM-GPR), applied to seven SREs across 38 stations in Ghana and Zambia, aimed at assessing their performance in rainfall detection and intensity estimation. Results indicate that the ENACTS product, which uniquely integrates a large number of station records, was the most corrigible SRE; in Zambia, nearly all BC methods successfully reduced the mean error on daily rainfall amounts at over 70% of stations. However, this performance requires further validation at independent stations not incorporated into the ENACTS product. Overall, the statistical methods (QM and LOCI) generally outperformed other techniques, although QM exhibited a tendency to inflate rainfall values. All corrected SREs demonstrated a high capability for detecting dry days (POD $ge$ 0.80), suggesting their potential utility for drought applications. A critical limitation persisted, however, as most SREs and BC methods consistently failed to improve the detection of heavy and violent rainfall events (POD $leq$ 0.2), highlighting a crucial area for future research.

Problem

• Correcting biases in satellite rainfall data occurrence and intensity
• Evaluating statistical and machine learning bias correction methods
• Addressing poor detection of heavy rainfall events after correction

Innovation

• Evaluated statistical and machine learning bias correction methods
• Hybrid techniques combined statistical and machine learning approaches
• Corrected satellite data effectively detected dry days but not heavy rainfall