Choosing a rainfall dataset can change the apparent long-term trend
A comparison of 19 global or near-global precipitation datasets finds that disagreement between products can rival the trends being measured. Satellite changes leave discontinuities, and rain-gauge correction does not consistently remove them. These differences complicate estimates used for water planning, agricultural risk and comparisons with climate models.
Science··Morning
Dataset differences rivalled the estimated change
Comparing 19 global or near-global precipitation products, researchers found that the choice of dataset can introduce uncertainty comparable to the trend being estimated. The peer-reviewed study evaluated annual mean precipitation, daily maxima and wet-day frequency.[1], [2]
Products without rain-gauge correction matched the reference trend’s direction on average only 58–64 per cent of the time. Across all products, the interquartile spread of trends was about 1.1–1.9 times the reference trend’s magnitude. Different datasets could therefore yield substantially different accounts of long-term rainfall change.[1]
Instrument transitions left jumps in long series
The common comparison covered 2001–2025, while the longer climate-model comparison extended across 1983–2025. Discontinuities in precipitation estimates often coincided with changes in satellites and observing systems. Rain-gauge correction did not consistently eliminate these breaks. Some corrected products also shared stations with the reference used to assess them, meaning their apparent agreement was not fully independent.[1]
Sparse station networks complicate comparisons
Reference measurements have their own constraints, including wind-induced undercatch and the difference between rainfall at one point and a wider grid-cell average. Uncertainty is particularly consequential where station networks are sparse. Averaging products does not remove every error because they can share inputs and processing assumptions. The findings concern how precisely rainfall trends can be measured, with consequences for infrastructure design and agricultural risk estimates.[1]
