Air temperature at 47 m

Half-hourly air temperature at 47 m on the CH-LAE tower, 2004-2025, gap-filled to completeness. The file is 02_METEO_TA_GAPFILLED_2004-2025 (parquet and CSV): 385,728 records on a continuous 30-minute middle-timestamp index (named TIMESTAMP_MIDDLE in the file) in local time (UTC+1, no daylight saving), from 2004-01-01 00:15 to 2025-12-31 23:45.

The interactive dashboard summarises the product on one page: coverage and provenance, seasonality, distributions, extremes and trends, with a table view behind every chart. It is standalone and works offline.

Method, evidence and checks:

02_METEO_TA — builds the product: screening, gap-filling, and the 2016 correction.
TA_HOMOGENIZATION_OPTIONS — the candidate corrections for the 2016 break, and why this one was chosen.
METEO_TA overview — the finished product year by year, with the residual steps recomputed on each run.

Which column to use

The measured column is not homogeneous across 21 January 2016, when the sensor and its acquisition system were replaced together.

Inside one sensor era, use TA_T1_47_1_gfXG, which is what the instrument reported.
Across 21 January 2016 (period means, trends, year rankings, anomalies, daily maxima, threshold-day counts) use TA_T1_47_1_HOMOGENIZED_gfXG. The step exceeds the climate signal over this record.

Columns

Both value columns are complete and both flags are defined at every record. The two value columns share the same flags, since the correction changes values, not provenance.

Table 1: Columns of 02_METEO_TA_GAPFILLED_2004-2025.

column	unit	description
`TA_T1_47_1_gfXG`	°C	Air temperature as the instrument reported it, gaps filled. Inhomogeneous across 2016-01-21.
`FLAG_TA_T1_47_1_ISFILLED`	-	Whether the value was measured, and if not, which method produced it.
`TA_T1_47_1_HOMOGENIZED_gfXG`	°C	The same series corrected so that the two sensor eras can be compared.
`FLAG_TA_T1_47_1_SOURCE`	-	Which sensor generation stands behind the record.

`FLAG_TA_T1_47_1_ISFILLED`

Filter on == 0 for measured records. The series is complete, so every other code marks a modelled value. Code 3 is not used.

Table 2: Fill codes and their record counts.

code	meaning	records	share
0	observed	369,065	95.7 %
1	gap-filled by XGBoost	14,912	3.9 %
2	gap-filled from timestamp information only (fallback)	43	0.01 %
4	linear interpolation across a short gap	178	0.05 %
5	reconstructed from NABEL at 49 m on the same tower	1,530	0.4 %

Codes above 0 account for 4.3 % of the product. The code-5 records are almost all January 2004 (1,488 records), plus one in February 2004 and 41 in July 2012.

`FLAG_TA_T1_47_1_SOURCE`

Table 3: Sensor-era codes, their record counts and the period each covers.

code	sensor era	period	records
0	Campbell CS215 on SDI-12	2016-01-21 14:15 to 2025-12-31 23:45	174,356
1	Rotronic MP101A, single-ended analog	2004-09-20 10:45 to 2015-12-31 23:45	197,739
2	acquisition changeover, sensor era undetermined	2016-01-01 00:15 to 2016-01-21 13:45	988
3	before the tower record begins	2004-01-01 00:15 to 2004-09-20 10:15	12,645

Codes 2 and 3 mark records that no instrument on this tower stands behind. Select FLAG_TA_T1_47_1_SOURCE <= 1 where provenance has to be certain.

The 2016 sensor change

On 21 January 2016 a Rotronic MP101A, read as an analog voltage, was replaced by a Campbell CS215 read over a digital line, together with the logger that read it. Two things changed at once, and TA_T1_47_1_HOMOGENIZED_gfXG corrects both.

The old chain read too cold, day and night. It carried a constant zero-point error of about -11 mV, which at 10 mV per °C left it reading roughly 1.1 °C below an aspirated reference on the same tower; the new sensor reads about 0.2 °C above that reference. The gap between the two eras is +1.3197 °C, added to every record before the changeover.

The two sensors warm differently in sunshine. Both sit in passive shields, which heat up in sun, and the newer one heats more. This part of the step appears only in daylight, so no constant can remove it. It was measured against NABEL, an aspirated fan-ventilated sensor at 49 m on the same tower, and subtracted from the later era. It is zero at night and reaches -0.59 °C at most.

Both corrections were fitted against NABEL and then scored against MeteoSwiss Lägern, a station 2.5 km away that entered neither of them:

Table 4: The step across the sensor change, measured against MeteoSwiss.

window	as measured	after both corrections
night	+1.31 °C	-0.01 °C
day	+1.69 °C	+0.06 °C
all hours	+1.51 °C	+0.03 °C

Over the whole record the correction more than halves the apparent warming. The difference between the 2005-2015 and 2017-2025 period means falls from +2.39 °C to +0.91 °C, and the Theil-Sen trend over annual means from +1.76 to +0.76 °C decade^-1. Those figures come from the complete columns; computed from measured records only they differ by less than 0.1 °C. Notebook 02 reports both.

Known limitations

A small difference in the shape of the day remains. The shield correction is fitted on the three years NABEL overlaps the newer sensor and applied over ten, so it is an extrapolation after 2018. Measured against MeteoSwiss on 2013-2015 against 2016-2018, the homogenised column still carries a daily-maximum step of +0.21 °C, a diurnal-range step of +0.18 °C and a daily-minimum step of +0.03 °C; over the full eras all three are smaller. The same statistics on the measured column step by up to +1.90 °C. Daily maxima and diurnal ranges are the quantities most exposed to what remains.
The record is made self-consistent, not absolutely accurate. The later era is standardised onto the earlier one, so the earlier era’s own warm bias in sunshine (about +0.4 °C at strong radiation, relative to an aspirated sensor) remains in both. What is removed is the change in that bias at the sensor swap, not the bias itself.
The records reconstructed from NABEL carry too little scatter. The 1,530 code-5 values come from a line fitted on NABEL, so they sit approximately 0.05 °C from it where a real measurement sits approximately 0.25 °C from it. Their means are sound; variance, extremes and any statistic of spread computed over January 2004 are understated.
2004 is 72 % modelled. The tower record begins 2004-09-20 10:45. January 2004 is reconstructed from NABEL rather than gap-filled, because the MeteoSwiss reference does not begin until 1 February 2004. The year carries source code 3 throughout its first nine months and is the least constrained in the record.