Summary

An improved method for merging weather station temperature data leads to revised temperature trends for the period 1900-2023 for the 6 largest southern provinces of Canada, compared to those I previously posted here and here. The general conclusions remain the same, but the details change somewhat. Because of the improved methodology, this post supersedes those posts. The main conclusions for the period 1900-2023 are:

1. Southern Canada daily high summer temperatures [Tmax] have warmed at only 1/3 the rate of daily low temperatures [Tmin]: +0.06 and +0.18 C/decade, respectively;
2. Averaged across southern Canada, 2021 and 1961 are the two hottest summers for daily high temperatures;
3. Even though 2021 is the hottest for Tmax, none of the individual 6 provinces has 2021 as record warmest; and
4. Tmax and Tmin trends are surprisingly uniform across those 6 provinces.

The “New” Methodology

In recent months I’ve spent a lot of time investigating various methods for combining different weather station records for the purpose of quantifying long-term temperature changes. One of the things I discovered is that, if there are few stations in a given year, doing a homogenization process such as the Menne and Williams (2009) Pairwise Homogenization Algorithm (PHA, used by NOAA and the BEST dataset providers) can lead to a random walk behavior as errors in the method in a single year will then propagate through all later years. This is probably not a problem in the U.S. since there are so many stations, but in other parts of the world it could be an issue. So, I think it is worthwhile to use an alternative methodology involving different assumptions.

After my previous posting of the aforementioned analyses of Canadian temperature trends, a few people (including John Christy) correctly pointed out that my straight averaging of all available stations in a province (or U.S. state) isn’t the best way to come up with a long term time series of temperatures. This is because as stations come and go over the years, they might be in different areas with different average weather. Of course, I already knew this, but ignored that nuance for the time being. But when I implemented a method that removes inter-station biases, I discovered that it did make some difference (as expected).

So, I implemented the merging procedure that John and I have used for many years with our UAH satellite temperature dataset, which is to remove relative station (or satellite) biases during overlap periods of time. This takes out any inter-station differences due to geographic location, altitude, urban heat island effects, poor siting of thermometers, equipment differences, etc. What isn’t accounted for is any spurious station temperature trend effects, say due to increasing urbanization, a sensor location or equipment change at that station, etc. So, for this initial version of the method I am assuming those changes average out over time. Of course, UHI effects would not average out over time since they almost always operate in one direction (spurious warming).

One issue is how to start the merging of stations together. I concluded that it is best to start with the stations with the longest periods of record, then add in the other stations, in ranked decreasing order of length-of-record, after removing each station’s offset (bias) relative to the average of all previously merged stations. As an example, the Alberta data merging (results shown later) involves a total of 950 separate stations, with varying lengths of record. The longest operating Alberta station was 111 out of 124 possible years (1900-2023). Half of the Alberta stations had periods of record of 15 years or less. The shortest period of record included was 2 years, because that is the minimum necessary to remove an average inter-station bias as well as have any time-variation information.

Results

The following figure shows the daily high (Tmax) and low (Tmin) summer (June-July-August) temperature trends, 1900-2023, for the 6 provinces from which I analyzed data. The other provinces have very few stations by comparison to these six.

There is a surprising (at least to me) level of consistency in the trends across the provinces. The Tmax trends average only 1/3 the Tmin trends, so summer nights are warming much faster than summer days. Again, urban heat island efects have not been removed, so it remains unknown how much of this difference is due to UHI effects, which are much more pronounced in Tmin than in Tmax.

The next figure shows the yearly time series averaged across those 6 provinces. I’ve included the linear trend as well as a 3rd order polynomial fit to the data, the latter to reveal the warmth during the Dust Bowl years of the 1930s.

Interestingly, even though none of the individual provinces had 2021 (the year of the epic late-June heat wave in western Canada) as the record warmest summer, the average across the 6 provinces did have 2021 as record warmest, barely edging out 1961:

Note that the Dust Bowl years of the 1930s shows up much more in the Tmax than Tmin data, probably due to lower humidity air. The cool summers of 1992-93 after the major eruption of Mt. Pinatubo also show up clearly.

Individual Province Time Series

Following are the 6 individual province time series of yearly summer Tmax and Tmin; all temperature scales span 8 deg. C of range for ease of comparison. I present these without comment, except to point out that the warmest BC year was 1958, not the epic heat wave year of 2021, the effects of which were maximized in this province. My next step after this is to apply the same methodology to the 48 contiguous U.S. states (CONUS), and compare to NOAA’s homogenized trends for those states.