Which rainfall errors can hydrologic models handle? Implications for using satellite‐derived products in sparsely gauged catchments

There is interest in applying satellite-derived rainfall products for water management in data-sparse areas. However, questions remain around how uncertainties in different products interact with hydrologic models to determine simulation skill. Most related work uses performance statistics that inherently combine rainfall magnitude, timing and persistence, making it unclear which product improvements should be prioritized. We applied six satellite-derived rainfall products in a conceptual hydrologic model (GR4J) across four Australian catchments with dense gauge data for comparison. We found that GR4J's inherent flexibility allowed it to filter errors in rainfall magnitude and variance through parameterization. Therefore, when rainfall observations for bias correction are unavailable, calibration of a flexible model could prove a useful alternative. However, the model was less able to compensate for errors in rainfall occurrence. In fact, the Probability of Detection score explained 59% of the variance in calibration performance (26% for validation), while overall bias explained just 14% (8% for validation). All products underestimated rainfall state persistence, but this had less influence on model skill. We then removed gauges from the observed data set to mimic data sparsity, finding that even a few gauges could reproduce rainfall occurrence and outperform satellite-derived products. Two data-sparse catchments in Vietnam were modeled to check whether the same learnings applied. The gauge data also performed best in Vietnam, and performance of most satellite-derived products was comparable to the Australian case. Efforts to increase the spatial and temporal resolution of satellite observations, which could improve rainfall detection, will enhance satellite-derived precipitation for hydrologic modeling.