Air temperature at 47 m

Half-hourly air temperature at 47 m on the CH-LAE tower, 2004 to September 2026, gap-filled to completeness. The file is 02_METEO_TA_GAPFILLED_2004-2026.09 (parquet and CSV): 398,832 records on a continuous 30-minute index TIMESTAMP_MIDDLE (middle of the half-hour, UTC+01:00, no daylight saving), from 2004-01-01 00:15 to 2026-09-30 23:45. 2026 is a partial year (January to September).

The interactive dashboard summarises the product on one page.

Method, evidence and checks:

02_METEO_TA: builds the product: screening, gap-filling, the 2016 correction, and the treatment of the May 2025 change and the June 2026 probe change.
TA_HOMOGENIZATION_OPTIONS: the candidate corrections for the 2016 break, and why this one was chosen.
METEO_TA overview: the product year by year. It still describes the earlier 2005-2025 version and does not yet show the 2026 changes.

Which column to use

The measured column is not homogeneous. The 47 m record comes from three probes, and the second of them changed state twice before it was replaced (see Sensor history).

Inside one probe state, use TA_T1_47_1_gfXG, which is what the probe reported.
Across 21 January 2016, 16 May 2025 or 5 June 2026 (period means, trends, year rankings, anomalies, daily maxima, threshold-day counts) use TA_T1_47_1_HOMOGENIZED_gfXG. The 2016 step alone is larger than the climate signal over this record.
Where provenance has to be certain, select FLAG_TA_T1_47_1_SOURCE in {0, 1, 4, 6}. Exclude code 7 from daytime statistics (daily maxima, diurnal ranges): it remains too warm in sun after homogenisation.

Columns

Both value columns are complete, and both flags are defined at every record and describe both value columns.

Table 1: Columns of 02_METEO_TA_GAPFILLED_2004-2026.09.

column	unit	description
`TA_T1_47_1_gfXG`	°C	Air temperature as the probe reported it, gaps filled. Not homogeneous across 2016-01-21, 2025-05-16 and 2026-06-05.
`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 with every probe state put on one level.
`FLAG_TA_T1_47_1_SOURCE`	-	Which probe, and which state of it, stands behind the record.

`FLAG_TA_T1_47_1_ISFILLED`

Filter on == 0 for measured records. Every other code marks a filled value. Code 3 is not used.

Table 2: Fill codes and their record counts.

code	meaning	records	share
0	observed	382,167	95.82 %
1	gap-filled by XGBoost	14,914	3.74 %
2	gap-filled from timestamp information only (fallback)	43	0.01 %
4	linear interpolation across a short gap (up to 1 h)	178	0.04 %
5	reconstructed from NABEL at 49 m on the same tower	1,530	0.38 %

Gap-filling is split into five periods, at each probe change and at the end of the NABEL record in 2018, so that no model is trained across a change of probe. Of the 1,530 code-5 records, 1,488 are in January 2004, one in February 2004 and 41 in July 2012.

`FLAG_TA_T1_47_1_SOURCE`

Table 3: Probe and probe-state codes, their record counts and the period each covers (TIMESTAMP_MIDDLE).

code	probe and state	period	records
0	Campbell CS215 on SDI-12, normal state	2016-01-21 14:15 to 2024-10-21 23:45	153,428
1	Rotronic MP101A, single-ended analog	2004-09-20 10:45 to 2015-12-31 23:45	197,739
2	acquisition changeover, probe 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
4	Campbell HygroVUE10 on SDI-12	2026-06-05 11:45 to 2026-09-30 23:45	5,641
5	probe handling or probe swap half-hour	2025-05-16 09:45 and 2026-06-05 11:15	2
6	Campbell CS215 after the handling of 16 May 2025	2025-05-16 10:15 to 2026-06-05 10:45	18,482
7	Campbell CS215 with excess daytime heating	2024-10-22 00:15 to 2025-05-16 09:15	9,907

No probe on this tower stands behind codes 2, 3 and 5. The measured value of the two code-5 half-hours is removed and modelled. To stay inside one probe state, select a single code.

Sensor history

The field name TA_T1_47_1 and its database tags did not change at any of these dates.

Rotronic MP101A, to 31 December 2015 (code 1). Read as a single-ended analog voltage by a CR10X logger. The logger programs name the probe and its conversion.
Campbell CS215, from 21 January 2016 (code 0). Read over SDI-12 by the CR1000 logger installed at the changeover, whose program names the probe. The 20 days between the two are an outage (code 2).
The CS215 with excess daytime heating, 22 October 2024 to 16 May 2025 (code 7). On 19 September 2024 the fieldbook records intake tubing attached at the 47 m sonic anemometer. In the following weeks the daytime excess of the 47 m probe over the 35 m sensor of the same tower grows beyond any earlier value; a rule on that excess dates the onset.
The CS215 after 16 May 2025 (code 6). The probe was handled during a site visit, and the daytime excess disappears. See The May 2025 change.
Campbell HygroVUE10, from 5 June 2026 (code 4). It replaced the CS215. See The June 2026 probe change.

Table 4: What TA_T1_47_1_HOMOGENIZED_gfXG applies to each source code.

source code	correction
1, 2, 3	the constant of the 2016 change, +1.3197 °C, added to every record
0, 7	the shield term of the CS215, subtracted: zero at night, at most 0.59 °C
6	a constant night term, +0.058 °C, subtracted at all hours; no shield term
4	the HygroVUE10 transfer onto the normal CS215, then the shield term of the CS215
5	as the period the half-hour belongs to (the handling record as code `7`, the swap record as code `6`)

The 2016 sensor change

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

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

The two probes warm differently in sunshine. Both sit in passive radiation shields, which heat up in sun, and the newer one heats more. This part of the step occurs only in daylight, so a constant cannot remove it. It was measured against NABEL, an aspirated sensor at 49 m on the same tower, and subtracted from the later era. It is zero at night and 0.59 °C at most.

Both corrections were fitted against NABEL and scored against MeteoSwiss Lägern (2.5 km away), which entered neither fit.

Table 5: The step across the sensor change, measured against MeteoSwiss, 2013-2015 against 2017-2019.

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

The correction raises the earlier era by the constant and lowers the sunlit records of the later eras by the shield term (by 0.15 °C on average over the 2017-2025 annual means of measured records). Over the whole record it more than halves the apparent warming. From the complete columns, the difference between the 2005-2015 and 2017-2025 period means falls from +2.39 °C to +0.92 °C, and the Theil-Sen trend over annual means from +1.76 to +0.77 °C decade^-1. From measured records only, the same figures are +2.47 °C and +1.00 °C, and +1.80 and +0.77 °C decade^-1. Notebook 02 reports both.

The May 2025 change

On 16 May 2025 the raw 1-minute record of the 47 m probe jumps by about 5 °C between 09:32 and 09:42, while the 35 m sensor of the same tower does not change: the probe was handled. The fieldbook records a site visit that day, with the heating connection of the tower inlet found melted and disconnected. The half-hour 09:30 to 10:00 is removed and modelled (source code 5).

The change is placed on the 47 m probe by two comparisons. The 35 m sensor against MeteoSwiss Lägern, which contains no 47 m measurement, changes by an ordinary amount at the handling (-0.36 °C above 500 W m^-2, 14 days after against 14 days before). The 47 m probe against the 35 m sensor changes in sun by -1.11 °C, more than on any other day of the surrounding half-year.

Code 7, the heated state from 22 October 2024, receives the shield term of the normal CS215 and nothing more. In strong sun it remains up to 0.92 °C too warm after homogenisation.
Code 6, from 16 May 2025, no longer has the daytime excess of the normal CS215 and receives no shield term. Its night level is compared with the normal CS215 of 2022-2024 against Lägern, month by month over March to September, and the difference, +0.058 °C, is removed as a constant. Its uncertainty is 0.22 °C, the year-to-year scatter of the same quantity among the reference years. After the correction, code 6 lies within 0.13 °C of the normal CS215 at night and in every radiation class. In a month hold-out (each month corrected with the term derived from the others) it departs from the normal CS215 by at most 0.215 °C, against up to 0.57 °C between one reference year and the others in the same month.

The June 2026 probe change

On 5 June 2026 a Campbell HygroVUE10 (SDI-12) replaced the CS215 at 47 m. The raw 1-minute record has a gap from 11:10 to 11:31 that day. The fieldbook adds the HygroVUE10 at this location on 5 June and records a change of its SDI-12 configuration with a new logger program on 11 June. The residual against MeteoSwiss Lägern changes level from that day, most in sun. The swap half-hour 11:00 to 11:30 is removed and modelled (source code 5).

NABEL ended in 2018, so Lägern is the only reference across this change. The transfer is fitted on two season-matched samples: the HygroVUE10 from 12 June to 30 September 2026 (5 to 11 June, before the configuration change, enter no fit) and the normal CS215 (code 0) on the same calendar dates of 2022 to 2024. It has two terms:

A radiation term, zero in the dark. In sun the HygroVUE10 reads cooler than the normal CS215 by an amount that grows with radiation, to 0.98 °C at 600 W m^-2.
A night term, a constant. At night the HygroVUE10 reads 0.11 °C cooler than the mean of the CS215 summers, less than the scatter among those summers (0.19 °C).

After both terms the HygroVUE10 lies within 0.05 °C of the CS215 level at night and in every radiation class of the fitting sample. In a month hold-out (each month refitted without itself and predicted) the held-out months depart from the CS215 level by at most 0.18 °C, against up to 0.58 °C before the correction and up to 0.38 °C between one CS215 summer and the others in the same month. The shield term of the CS215 is then applied, so the HygroVUE10 era ends on the same reference level as the other eras.

Known limitations

A small difference in the shape of the day remains across 2016. The shield correction is fitted on the three years in which NABEL overlaps the CS215 (2016-2018) and applied until May 2025 and, through the transfer, to the HygroVUE10 era, so after 2018 it is an extrapolation. Against MeteoSwiss, 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; the measured column steps by up to +1.90 °C in the same statistics. Daily maxima and diurnal ranges are most affected by what remains.
The heated CS215 (code 7) is under-corrected in sun. Daytime values and daily maxima from 22 October 2024 to 16 May 2025 are biased warm, by up to 0.92 °C in strong sun. The excess grows gradually and its onset is dated by a rule over a month-long window, so records shortly before 22 October 2024 may already carry part of it.
The night term of code 6 is smaller than its uncertainty (+0.058 °C against 0.22 °C). It is applied because the change of probe state is established, not because the night difference is resolved, and it rests on one reference 2.5 km away.
The HygroVUE10 transfer rests on one summer and one reference. It is fitted on 12 June to 30 September 2026 against Lägern and applied as fitted to every HygroVUE10 record, including 5 to 11 June. Its behaviour in winter, under snow or rime and at low sun is untested. The CS215 level it is fitted to is a mean of three summers that differ among themselves at night by up to 0.19 °C, more than the night term of 0.11 °C. It is to be refitted once the HygroVUE10 record includes a winter.
The record is made self-consistent, not absolutely accurate. The later eras are standardised onto the shield behaviour of the MP101A, so its warm bias in sunshine (about +0.4 °C at strong radiation, relative to an aspirated sensor) remains in all of them.
Records reconstructed from NABEL have too little scatter. The 1,530 code-5 values come from a line fitted on NABEL, so they lie about 0.05 °C from it where a real measurement lies about 0.25 °C from it. Their means are sound; variance, extremes and other measures of spread over January 2004 are understated.
2004 is 72 % modelled. The tower record begins 2004-09-20 10:45. January 2004 is reconstructed from NABEL because the MeteoSwiss reference begins only on 1 February 2004. The first nine months of 2004 carry source code 3 and are the least constrained part of the record.
2026 is a partial year. A 2026 annual mean covers January to September only.