PPM_DB3 · climatol 4.5 · 250 stations · 2000–2025
Atacama Desert before and after climatol
The desert is not one regime but four, and treating them together was the mistake. Once separated, homogenization stops inventing breaks and — what matters most here — stops erasing the rain that did fall.
A single run, without separating regimes, flagged 44 breaks. Segmented by regime, 4 remain: Visviri, January 2005; El Corral, January 2010; Los Loros, January 2016; Rio Lluta En Alcerreca, January 2021. The difference is not leniency but the arithmetic of the reference: between stations of the same regime the median monthly correlation is 0.78; between different regimes, 0.16. Comparing the Altiplano with the coast manufactures the breaks that are later “corrected”.
That so few remain depends on a decision worth stating. The two-step design says that the annual scale detects the breaks and the monthly scale corrects them. climatol disables monthly detection when it receives metadata, so if the metadata are passed only when the annual step found something, some regimes search and others do not. Here they are always passed, empty if necessary: all four regimes receive the same treatment. Under the previous criterion six more monthly breaks appeared; they were artifacts of the procedure, not of the data.
The Cancosa break deserves separate attention: climatol placed it in March 2019 without being told anything, within two months of the date at which the cleaning had already truncated that series because it was no longer credible. On closer inspection it turned out to be not an inhomogeneity but a change in the measured variable: up to 2018 it concentrates 96 % of its rain in December–March, like any Altiplano station; from 2019 its maxima are July (446 mm), June (432) and May (421). And the values do not come from the primary portal but from CR2, the redistributor. The series is kept flagged, not deleted.
end up as zero (previously 110)
(the single run gave 44)
is observed
run separately
The six tests
One station per locality
Annual totals. The wide gray bar is the homogenized and infilled year; the colored core is what the rain gauge measured. If the core reaches above the gray, homogenization lowered that year; if the gray shows above it, homogenization raised it. The observed year is drawn only when all twelve months were measured —a month without data is not a dry month—so its absence marks the years that exist only thanks to infilling.
Month by month: measured above, homogenized below
Gray bands are months without observation. In the ORIGINAL row the line disappears there; in HOMOG. it continues, and that continuation is an estimate, not a measurement. On the right, the mean annual total of each version.
The decisive validation. Two rain gauges from different institutions less than 1 km apart measure the same rain, and neither was declared to climatol as the other's pair: if homogenization damaged anything, it would show here. There are 15 pairs with at least twelve months measured at both.
Over the months both measured, nothing changes. The RMSE between pairs is 11.33 mm before and 11.33 mm after, and the correlation is 0.891 in both cases: homogenization does not touch the measured cell. What does change is which months are compared, because the homogenized series is complete: it adds 2,409 month pairs in which one of the two is an estimate, and there the RMSE is 9.78 mm. An earlier version of this figure read that difference as “infilling degrades the neighbors”; it was the arithmetic of different sets of months.
What v5 fixed: not infilling what cannot be infilled. In v4, Río Copiapó en Pastillo measured 278.4 mm in March 2014, and infilling gave Altar de la Virgen, 522 m away, 0.0. The first attempt was to reassign it its neighbor's regime: that month rose to 227.2 mm, but others went the opposite way —307 mm where Pastillo measured 81. climatol infills by scaling with each series' own mean, and Altar's mean came from four years, 2022-2026, with a July of 188 mm: any such mean biases every month it infills. Since v5, a series with fewer than 48 measured months is neither homogenized nor infilled: it remains in the observed database, flagged, and 25 series that used to be infilled drop out.
And what it did not fix. In the remaining pairs, infilling behaves almost as in v4: the overall improvement comes from not infilling where it could not be done, not from infilling better. The pairs that diverge most are still in the Altiplano —Putre, Alcérreca, Chilcaya, Tacora—where a convective storm falls on one rain gauge and not on the one next to it, and no regional estimate can reproduce it. For spatial statistics or comparisons between stations, use only bandera = 0.
What this work does not solve. 39.0 % of the matrix is infilled. It serves to detect breaks, which was the goal. To compute trends, or for any comparison between stations, use only the observed cells.