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

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.