Arctic temperature trends and data sparsity

Whilst browsing the wilder shores of skepticism (well, I'd just been to Ikea and needed some light relief...) I came across the inaccurately titled "Reality in Arctic temperature trends. Scroll down about 1/4 of the way to the 1880-2004 temperature plot. So... temperatures higher in 1935-1945 than now? Interesting! And using CRU data too. How come... Well, one funny thing is that he calls this "A sobering dose of reality" - presumably forgetting that elsewhere he has attacked the Jones data as the spawn of the Devil. A second funny thing is that he is using [70,90]... [60,90] is more usual. Would you get the same results for [60,90]? And are there really many stations between [70,90] in the early period?

I'm sure you can guess the answers, and its "no" to both. Have a look at my pic (but be careful, there are lots of lines...). The top graph is [70,90]. The bottom graph is [60,90]. Both show the area-averaged temperature anomaly (in black; the 13-month running mean is in blue) from the HadCRUT2v dataset, in 100's of oC, which is why the left hand scale is 100 times bigger than you think it ought to be. Both plots have the same general shape, but for the wider area the current (last 10 years, mean given by red bar) temps are higher than for the 1935-1945 average. But even for [70,90] the temps in 1935-45 are only marginally higher than now - about 0.1 oC - hardly "much higher than today" as our septic claims.

But... look at the green lines. The lower green line on each plot is the fraction of the area covered by obs, on the right-hand scale. So for [60,90] about 40% of the area is observed, since 1960. In the 1940's, about 30%. For [70,90] about 20% is observed, recently (though with a huge annual cycle: far more people about in summer!) and less than 10% in the 1940's. Our septic complains that the "Arctic Climate Impact Assessment (ACIA) start their temperature records in 1960". Errm yes, well that might well be a good idea. Perhaps the ACIA people actually bothered to look at the data rather than just area-averaging it.

In fact, to my not-great-surprise, the ACIA people do indeed look at temperatures before 1960 (hint: if a septic sez something is true, its probably false...) and even draw nice maps of the trends at various time intervals: see the ACIA sci report, p36 and after. But they note the data sparsity problems early on.

The total *number* of filled 5x5 degree gridboxes is the upper green line on each plot, and the scale is (conveniently) the [0,400] of the upper half of the temperature scale (has your mind exploded yet?) *except* that for [70,90] that would be too small to see so I've multiplied it by 10 (boom!). So at the time of that huge (and rather suspicious...) jump in the upper plot at 1919, there were only 5 (=50/10) stations. For [60,90] there are nearly 200 filled boxes, recently.

Just looking at fraction-of-obs can be a bit dry, so here are maps of gridboxes filled (with their anomaly values, no in sensible units) for July 1919, 1940 and 2000. Note that using July maximises the filled boxes for the year. Its pretty obvious that 1919 is *very* sparse; 1940 is sparse; but even 2000 isn't exactly packed, north of 70; though its pretty good from 70 to 60 (oh, the black circles are 60 and 70 N, of course).

So... what do we learn from all this (apart from never trust the septics, but we knew that already)? We learn that plucking a dataset off the shelf and playing with it and only showing the end result may well mislead... we learn that you should be cautious with sparse data.