Incoming shortwave radiation at 47 m
Half-hourly incoming shortwave radiation at 47 m on the CH-LAE tower, 2004 to September 2026, gap-filled to completeness. The file is 01_METEO_SW_IN_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.
The interactive dashboard summarises the product on one page.
Method, evidence and checks:
01_METEO_SW_IN: builds the product (merge of the two screenings, 2012 corrections, nighttime offset, gap-filling) and shows that the record is homogeneous across both hardware changes.RADIATION_SENSOR_CONTINUITY: compares four radiation sensors and attributes each level change in the record to one instrument.- Meteo Product Chain:
TAandPPFD_INuse this product as a gap-filling driver.
This series needs no homogenisation. The same sensor was read through the same calibration across the January 2016 acquisition change, and the December 2021 replacement of the radiometer did not move the level either. There is one value column and no _HOMOGENIZED column. A slow departure of the sensor from its references from 2013 is described under Known limitations.
Columns
Every record carries a value, and the flag is defined at every record.
01_METEO_SW_IN_GAPFILLED_2004-2026.09.
| column | unit | description |
|---|---|---|
SW_IN_T1_47_1_gfXG |
W m-2 | Incoming shortwave radiation, gaps filled. Complete, non-negative, exactly zero at night. |
FLAG_SW_IN_T1_47_1_ISFILLED |
- | Whether the value was measured, and if not, which method produced it. |
FLAG_SW_IN_T1_47_1_ISFILLED
Filter on == 0 for measured records. Every other code marks a filled value. Code 4 is not used.
| code | meaning | records | share |
|---|---|---|---|
| 0 | observed | 364,830 | 91.5 % |
| 1 | daytime gap, filled by the XGBoost model | 17,411 | 4.4 % |
| 2 | daytime gap, filled by the timestamp-only fallback model | 697 | 0.2 % |
| 3 | nighttime gap, set to zero | 15,894 | 4.0 % |
Code 3 is not an estimate: the sun is below the horizon and the value is zero. Codes 0 and 3 together are 95.5 % of the record.
Coverage
The tower measurement begins on 2005-09-14 11:15, when the radiometer was wired to the tower logger. Everything before that is modelled.
| year | measured | why |
|---|---|---|
| 2004 | 0 % | before the radiometer was connected to this logger |
| 2005 | 30 % | the record begins on 14 September |
| 2012 | 93 % | logger clock error, power-supply failure and storm damage |
| 2016 | 94 % | the January outage during which the logger was replaced |
In 2012 a logger clock error shifted one block of August by 15.5 hours, and a tower power-supply failure in late July and August and storm damage in late October and November left records that could not be repaired. The clock error was corrected, the other two periods were removed, and all of it was gap-filled.
How the gaps were filled
Daytime gaps are filled by an XGBoost model, trained separately for 2004-2018 and 2019-2026. The split is where the NABEL sensor at 49 m on the same tower ends, so that no model is trained on one driver set and applied to the other. Until 2018 the drivers are NABEL and MeteoSwiss Lägern (2.5 km away); from 2019 only Lägern remains, so the later fills rest on the more distant sensor. Code 2 marks the few half-hours where no driver was available. Drivers enter as measured, never gap-filled. Nighttime gaps are set to zero. A reference is used only as a driver and never replaces a measured value.
Known limitations
- The sensor reads about 3 % high relative to its references from 2013. NABEL at 49 m, MeteoSwiss Lägern and the co-located PAR sensor all show it at once, which places the change on this radiometer. It develops over about three years rather than at a date, and the fieldbook records no maintenance or calibration of this radiometer between its installation in 2005 and its removal in 2021. The change is within the field uncertainty of a pyranometer of this class and is not corrected, because there is neither a boundary at which to apply a correction nor a known fault to justify one. Allow for a few per cent when comparing the middle of the record with its beginning.
- The MeteoSwiss gap-filling driver changed level in October 2010.
SW_IN_LAE_MS, the only driver after 2018, steps by about 5 % on 6 October 2010, when that station’s radiation instrumentation was rebuilt. Both tower radiation sensors and NABEL step against it together. The fills are not affected, because a driver supplies the state of the sky, which a scale change does not alter. Do not read a difference between this product and MeteoSwiss Lägern as evidence about the tower across that date. - Nighttime values are exactly zero, not measurements. The small drifting nighttime offset of the pyranometer was removed per day before gap-filling. For instrument noise or the offset itself, use the screened database series.
- 2004 and most of 2005 are entirely modelled. They precede the measurement and rest on MeteoSwiss Lägern alone, which itself begins on 1 February 2004. These are the least constrained parts of the record.
The hardware changes of 2016 and 2021
The January 2016 acquisition change did nothing measurable to this series. It is recorded here because the same changeover moved TA_T1_47_1 by 1.3 °C. For TA the sensor and its conversion changed; for shortwave radiation neither did. Every surviving logger program, three CR10X programs of 2005-2006 and the CR1000 program installed at the changeover, reads the same Kipp & Zonen CNR1 (serial number 020484) on a 25 mV differential channel at multiplier 99.7009 (the reciprocal of its sensitivity) and offset 0. Across 2016 every ratio among the three shortwave sensors changes by less than in an ordinary year. The largest change, 1.4 %, is in the pair that does not involve the tower. Notebook 01 asserts this on every run.
The December 2021 replacement of the radiometer did not move the level either. A CNR4 replaced the CNR1, with a sensitivity of 13.89 instead of 10.03 µV/W/m². The date coincides with the change of screening software, so the two possible causes of a step cannot be separated, but neither produces one.