Do two rainfall products agree, year to year?

Notes

• Totals are computed from monthly data on the 0.5° grid of the main analysis; a year counts only if every month of the period is present. 3-month periods crossing the year end (NDJ, DJF) are labelled by the year they end in.
• Rainy season: the calendar months whose GPCC 2001–2020 climatology is at least the monthly mean, summed within a hydrological year that starts the month after the driest month (so a November–March season counts as one season), and labelled by the calendar year of its last rainy month. The months used for the selected pixel are shown under the scatter.
• Admin-1 (Natural Earth, not COD boundaries) and country values are area-weighted means of the 0.5° cells assigned to the unit by majority, using only cells where both products have data that year. Small units may have one or no cell.
• Values are stored at 8-bit precision per pixel: typically 1–2 mm on annual totals, up to ~50 mm in the wettest pixels — negligible for correlation, visible only in the last digits of the scatter values.
• Physically implausible months are masked in every product: above both 1,500 mm and 15× the cross-product climatology for that pixel and month, and more than 5× the median of all other products (379 pixel-months in all — source errors such as CPC's 21,000 mm over Italy in February 2026 or IMERG Late's 10,000 mm over Siberia in January).
• IMERG Late October–November 2024 is excluded (NASA calibration incident). CHIRPS/CHIRP v2 cover 50°S–50°N, v3 60°S–60°N. The JRC ASAP blend is CHIRPS v2 within ±50° and ERA5 beyond, so it is not listed separately.
• "Detrended" removes each product's own linear trend before correlating, so shared trends (or opposite drifts) do not inflate or deflate the agreement on year-to-year variability.