Homogenizing a Warming Station Record
A regional climate center is homogenizing a station record for 1950–2020. Metadata: the station sat on a downtown rooftop until 1987, when it moved 9 km east to an airport; the liquid-in-glass shelter was replaced by an MMTS unit in 1996; observation time changed from 17:00 to 07:00 local in 2003. Gridded population for the cell grew from 12,000 to 180,000 over the record. The raw annual-mean series warms at 0.42 °C per decade, with step-like jumps at 1987 and again at 2003 that appear in Tmax but not in Tmin. The 1987–1995 segment has 14% missing daily values, and four years contain fewer than 300 valid days. A rural station 40 km away warms at 0.18 °C per decade, and a reanalysis series for the grid box warms at 0.25 °C per decade. The center must publish an adjusted trend with an uncertainty range and document each adjustment. The airport site now sits 400 m from a taxiway that was extended in 2011.
1. Rank the known discontinuities by likely influence on the annual-mean series, state the expected sign of each, and explain why the jumps appear in Tmax but not Tmin.
Read the worked answer
Ranking
- 1987 rooftop-to-airport relocation (largest). A downtown rooftop site is warm-biased, mainly through reduced nocturnal radiative loss and building heat in the minimum, plus wall re-radiation in the maximum. Moving to open airport exposure almost always produces a cooling step.
- 1996 liquid-in-glass to MMTS. The smaller, better-ventilated MMTS shelter produces a documented cool bias in Tmax of a few tenths of a degree. Correcting it means nudging pre-1996 Tmax down or post-1996 Tmax up by a site-specific offset.
- 2003 observation time, 17:00 to 07:00. Observation time determines which part of the diurnal cycle is captured in the climatological day, so a change systematically alters Tmax, Tmin, and diurnal range. The effect is typically a few tenths of a degree, and it is exactly the kind of bias that aliases into the trend if left uncorrected.
- Urban growth, 12,000 to 180,000. Raises Tmin far more than Tmax (nocturnal urban heat island), so it contaminates the minimum series over the whole record rather than producing a step.
- 2011 taxiway extension. Small, local, and recent; treat as an exposure note.
Why the jumps are in Tmax Exposure and shelter changes (rooftop, MMTS) affect the daytime maximum most directly. Urbanisation and the observation-time change act differently: an urban signal would show first in Tmin. The fact that the steps appear in Tmax and not Tmin points to instrumentation and exposure as the break causes, with the urban signal carried by the long-term Tmin trend.
2. The record is 14% incomplete with four years below 300 valid days. How does that affect the trend, and what is your handling rule?
Read the worked answer
The problem An annual mean computed from 300 days is a noisy estimate of the true annual mean, and the noise is not random: missing days usually come from equipment failures, staffing gaps, or automation transitions, which tend to cluster in particular seasons. If the 1987–1995 segment is systematically missing one season, its annual means are biased relative to complete years, and because the missingness is not spread evenly along the record, that bias produces a spurious step in the trend.
Handling rule
- Set a completeness threshold for a valid annual mean - commonly at least 10 of 12 monthly means adequate, or roughly 330 valid days - and flag years that fail rather than publishing them unadjusted.
- For flagged years, either exclude them from the trend fit or fill them from monthly anomalies regressed against reference stations; do not fill with a raw monthly climatology, which re-inserts the bias.
- Report how many years were excluded and show the trend computed both with and without gap filling. If the two differ materially, the missing data are not random and that must be disclosed.
- Apply the same rule to the 1987 breakpoint analysis, so the breakpoint magnitude is not inflated by a cluster of incomplete years right after the move - which is exactly where this record's 14% gap sits.
3. Estimate the non-climatic contamination and propose a defensible published trend with uncertainty. What would you tell a journalist asking whether the town has warmed?
Read the worked answer
Contamination estimate The rural station (0.18 °C/decade) and the reanalysis box (0.25 °C/decade) bracket the regional signal at 0.18–0.25 °C/decade. The raw station value of 0.42 °C/decade leaves 0.17–0.24 °C/decade unaccounted for, which is the plausible size of the rooftop, shelter, observation-time, and urban biases. A 0.42 °C/decade over 70 years equals about 2.9 °C total warming, which is implausibly large for this region and is itself a warning that the raw series is contaminated.
Do not simply subtract the rural trend. The two sites differ in elevation, exposure, and their own history, and the rural station may carry its own breaks. The defensible procedure is objective breakpoint detection against a reference series of neighbouring stations, adjustment of each detected break, and then a check that the adjusted series converges toward the 0.18–0.25 °C/decade envelope.
Publish something like 0.2–0.3 °C/decade, with a stated uncertainty derived from the spread across homogenization methods, and a full table of the adjustments applied.
For the journalist The area has warmed, but the station's raw number overstates it by roughly a factor of two. The corrected local trend is broadly consistent with the surrounding region. Note also that minimum temperatures have risen more than maximums, which is consistent with both the regional signal and a modest urban effect - the adjustment reduces the urban contribution but does not erase the warming. Distinguish a single station from a regional average.