Monthly Precipitation of the Atacama
ESEN

How it was made · 9 steps · 2000–2025

PPM Atacama: from measurement to publication

The path each precipitation value followed, from the network that measured it to this page: what was done to it at each step, which checks it had to pass and which series were left out.

View as interactive diagram best viewed on a large screen

Measurement

Twelve networks, one quality control and one clean database: flag instead of delete, and the measured zero is data.

  1. 1

    Sources

    12 networks

  2. 2

    Quality control

    Estévez et al. (2011)

    flag, don't delete
  3. 3

    PPM_DB3

    consolidated

    0 mm is data

The networks and their stations

NetworkCountryWith dataIn climatol
DGAChile156144
DMCChile6231
SENAMHIPeru2424
SENAMHIBolivia1713
INIAChile1615
CEAZAChile1411
APECO-FDFChile83
GHCNseveral76
INDAP-CONADIChile20
SQMChile22
MMA-DMCChile10
SNIHArgentina11
Total310250

“In climatol”: the stations with at least 48 measured months, which are homogenized.

Quality control

The record-structure check and the tests of Estévez et al. (2011) for weather-station networks, with fixed and dynamic range counted separately. Four apply to monthly rainfall totals, two do not, and one applies in part.

Before the tests: each month has to be complete

≥ 80 %
of the data · CEAZA, INIA, SNIH and the local archive
≥ 25 days
SENAMHI Peru
> 20 days
DGA, according to its own flag
no day count
DMC, CR2 and GHCN publish the monthly total without stating how many days it covers
TestWhat it checksApplies?How it is applied
Record structurecodes, dates and complete files; a failure is an errorappliesmissing-value codes and negatives → missing data · duplicates removed · fragments of the same site joined · coordinates corrected where two DGA records agree
Fixed rangephysical limits; a failure is an error and goes no furtherappliesimpossible records are removed: stations that do not measure rainfall and obvious regime breaks
Dynamic rangeagainst what is plausible at that site; a failure is suspectappliesmore than 6 times the neighbors' maximum (≤ 60 km) that month: flagged, not deleted
Stepthe change between two consecutive measurementsdoes not applyrainfall is discontinuous: Estévez does not apply it
Internal consistencyvariables of the same station against each otherdoes not applythere is a single variable
Persistencea sensor stuck on one valuein parta whole year at 0.0 mm at a station that usually records rain is flagged for review
Spatial consistencyagainst the neighboring stationsappliesa zero becomes missing data if its neighbors at a similar elevation measured more than 20 mm · then, climatol

Also

DGA > CR2
the DGA portal takes precedence month by month; CR2 only fills what is missing
s/p = 0 mm
at the DMC; a “.” is missing data
DEM
elevation comes from the DEM where it is missing, is 0 or does not match the terrain
± 500 m
each station is compared with its neighbors at a similar elevation: OK, suspect zeros, bias or not assessable

Estévez, J., Gavilán, P. and Giráldez, J. V. (2011). Guidelines on validation procedures for meteorological data from automatic weather stations. Journal of Hydrology, 402, 144–154.

Regime and inheritance

Under 4 years is not infilled. Reliable series estimate their regime; those without data in every month inherit it.

  1. 4

    Regime

    only series of 4+ years

    reliable ≥3 obs/month
    outExcluded < 4 measured years
    feeds · not reliableInherits regime from reliable neighborsthen on to climatol

How the regime is assigned

A
Altiplano: 60% or more of the rain falls between December and March
B
coast: 25% or less in summer; it rains in winter
C
transition: in between
D
hyperarid: less than 5 mm a year

A series is reliable with 48 measured months and 3 values in every calendar month. Those that have the months but not the values in every month inherit the regime of their reliable neighbors (≤ 60 km, ≤ 800 m elevation difference).

climatol in two steps

The exploratory run sets inht; the annual run detects the breaks and the monthly run corrects them.

    .dat / .est
  1. 5

    Exploratory

    sets inht

    in · references *ERA5-Land
    2. inht
  2. 6

    climatol

    detects and corrects

    year = 12 months
  3. 7

    Database v5

    validates

    check

How it is run

4 runs
one per regime: each series is compared only with those of the same climate
ERA5-Land
its pixels, at the station locations, are the reference series
inht
the break-detection threshold comes from each regime's exploratory run
annual → monthly
the annual total (only years with 12 measured months) detects the breaks; the monthly series is corrected and gap-filled

Publication

A SQLite / D1 database, with CHIRPS at each station, that rebuilds the figures and serves the API; rasters are downloaded at /rasters.

  1. 8

    SQLite · D1

    13 tables

    inCHIRPS at each station
  2. 9

    Site + API

    atacama.gfuentes.cl

    API · /rasters
    in · Zarr in R2Downscaling to 7.5″ elevation + NDVI
    from: GWR + CHIRPS

How it is published

SQLite · D1
the same database for download and for the API
CSV
the series and the stations, to open in a spreadsheet
API
stations and series by query
versions
every version is kept; the previous one is marked as superseded

Rasters

GWR per month with three safeguards, compared with ANUSPLIN, CR2MET and CHIRPS; downscaling to 7.5″. Only the Zarr goes to R2.

  1. Comparison ANUSPLIN · CR2MET · CHIRPS
    from: GWR + CHIRPS · LOOCV + exclusion
  2. Downscaling to 7.5″ elevation + NDVI
    from: GWR + CHIRPS

How they are made

GWR per month
ln(PPM + 1) ~ ln(CHIRPS + 1), with the bandwidth chosen by cross-validation
3 safeguards
minimum slope, CHIRPS within the neighbors' range and a cap of 1.5 times their maximum
5 km → 7.5″
each cell is split according to elevation and NDVI
25–100 km
exclusion radii used for validation, against ANUSPLIN, CR2MET and CHIRPS

Inputs: CHIRPS v2.0, MODIS NDVI (MOD13Q1), Copernicus DEM GLO-30 and, for comparison, CR2MET v2.5.

In summary

Decisions

  • Month without data ≠ dry month
  • Flag, don't delete
  • climatol only changes measured values where it finds a break

Checks

  • Minimum of 48 months to infill
  • inht only from the exploratory run
  • Before/after, same months

Raster validation

  • LOOCV and 25–100 km exclusion
  • ANUSPLIN: same model, same cells
  • Plausibility: 1.5 × the measured maximum