Botswana illustrates some of the challenges in analysing temperature data in Africa. Simply, there just isn't enough data. In fact there is more temperature data for California than there is for the whole of Africa.
In total, there have only ever been twenty weather stations in Botswana. Of these only six are medium stations with more than 480 months of data. This includes only two with more than 900 months of data and only three with any data before 1960. There are no long stations with over 1200 months of data and there is no data before 1900. Moreover, as the map in Fig. 37.1 illustrates, those medium stations that do exist are not evenly distributed, but are instead concentrated along the South African border.
Fig. 38.1: The temperature trend for Botswana since 1917. The best fit is applied to the interval 1917-1976 and has a negative gradient of -0.60 ± 0.90 °C per century. The monthly temperature changes are defined relative to the 1991-2010 monthly averages.
The lack of data also means that the overall temperature trend is very sensitive to the individual contributions from one or two atypical station records. This is highlighted in the difference between the trends shown in Fig. 38.1 above and Fig 38.3 below.
The trend in Fig. 38.1 was constructed by the usual method of averaging the temperature anomalies from the various stations for each month from the earliest temperature observation (which for Botswana is January 1917) until the latest (October 2013). As I have explained before, the monthly anomaly is the change in the monthly temperature from a pre-defined reference temperature for that month and they are used so that temperature changes over time for different stations and different regions may be more easily compared. The mathematics of their calculation is explained here. However, there are a number of problems that arise when trying to calculate these monthly reference temperatures (MRTs).
The first thing to note is that the MRTs are different for each station record, and are also different for each of the twelve calendar months within each record in order to eliminate, or at least minimize, seasonal variations. The MRTs are usually determined by averaging a set of temperature readings from the same calendar month within that particular temperature record (although some climate science groups appear to corrupt this process by using a process of homogenization to combine data from adjacent stations). Ideally this averaging is done by choosing a time interval that is both reasonably long, and also one over which there is very little overall change in temperature. For these reasons a thirty year time interval of 1951 to 1980 would probably be best. It is long enough for the MRT values to be close to the true mean, and it appears that many temperature records around the world exhibit much less variation in temperature over this time period in comparison to both earlier and later time intervals. It is also the time interval that most of the climate science groups initially chose when highlighting climate change in the 1980s and 1990s.
Unfortunately, in many countries in the Southern Hemisphere there is much less temperature data before 1960 compared to that which was recorded post-1980. For that reason it is often better to choose a later time interval such as 1961-1990, or a shorter one of perhaps only twenty years, say 1981-2000.
The next problem, though, is that the temperature records in a particular region or country will not all be of the same length. More importantly, they usually have different amounts of data within the the MRT interval. The question here is, how many months of data do you need to average in order for the MRT to be sufficiently accurate? The higher the proportion needed, the more station records that will be excluded. Ideally we would want all stations to have 100% data coverage within the MRT time interval for all twelve months of the year. But equally, we would, ideally, also want all the station records to be included in the overall trend. In practice very few stations would meet the criterion of 100% data coverage so a lower threshold needs to be set. I generally choose between 40% and 60% with a higher threshold for a shorter MRT interval.
Ultimately the only way to determine the optimum method of determining the MRTs is to test different approaches. In the case of countries with a large amount of data, the different choices for the MRT time interval and the data coverage threshold have little overall impact. However, for countries like Botswana with small numbers of stations, these choices matter because the exclusion of one or two sets of station data can have a major impact on the final temperature trend. This is illustrated in the difference between the trend in Fig. 38.1 above and the one in Fig. 38.3 below.
The temperature trend in Fig. 38.1 was constructed by first calculating the monthly reference temperatures (MRTs) for each station for the period 1991-2010. For this analysis only the fourteen stations with more than 180 months of data in total were included in the process (for a list see here). In addition, in order to optimize the accuracy in determining the MRT for each month for each station, only stations with data in more than 60% of months (i.e. 12 months) in the MRT period of 1991-2010 were included in the calculation. This resulted in twelve station records being included and two being excluded. The resulting number of station records incorporated in the trend for each month is shown below in Fig. 38.2.
Fig. 38.2: The number of sets of station data included each month in the temperature trend for Botswana when the MRT interval is 1991-2010.
Unfortunately, one of the stations that was excluded was Gaborone (Berkeley Earth ID: 152785), which is one of only three stations with any data before 1959. The other station excluded was Mahalapye (Berkeley Earth ID: 5699) which only has data from 1961 to 1990. The effect of including both these stations can be seen in Fig. 38.3 below. The effect is to change the temperature trend before 1976 from a negative trend of -0.60 °C per century to a positive one with a trend of +0.71 °C per century. This was achieved simply by changing the MRT interval to 1961-1990. While this resulted in the inclusion of the two stations at Gaborone and Mahalapye, it also meant that six stations with virtually no data before 1990 were excluded. This in turn has had a slight impact on the trend from 1990 onward, and in particular the magnitude of the cooling from 2002 onward.
Fig. 38.3: The temperature trend for Botswana since 1917. The best fit is applied to the interval 1917-1976 and has a negative gradient of +0.71 ± 0.41 °C per century. The monthly temperature changes are defined relative to the 1961-1990 monthly averages.
What the Botswana temperature data illustrates is the difficulty of deriving conclusive conclusions about climate change when there is insufficient data. The temperature trend before 1976 could be strongly positive (as shown in Fig. 38.3) or strongly negative (as shown in Fig. 38.1), depending on how representative the Gaborone data is of the country as a whole. Given previous evidence from Australia, Indonesia and South America regarding the disparity in temperature trends between large cities and the rest of the country, I would suggest that the Gaborone data is more likely to be an outlier even though Gaborone is hardly a megacity (its population is about 230,000). In which case it is more likely that the temperature trend in Botswana before 1976 would be very similar to that for South Africa (i.e. stable and flat) rather than the more or less continuous warming trend that has been claimed by groups such as Berkeley Earth (see Fig. 38.4 below).
Fig. 38.4: The temperature trend for Botswana since 1860 according to Berkeley Earth.
The other feature of note in both Fig. 38.1 and Fig. 38.3 is the large temperature rise from 1980 until 2002, followed by a smaller but significant decline. This temperature rise coincides with a much smaller one seen in the South Africa temperature data (see Fig. 37.2), but the warming in Botswana is about four times larger. It may be tempting to discount this warming as spurious or just bad data (as many climate scientists do when the data is not to their liking), but it features in too many different station temperature records to be ignored that easily. Instead it hints at the possibility of a more worrying phenomenon for climate scientists: namely that natural fluctuations in the regional temperature could be much larger and more persistent than they currently accept is possible.