Tuesday, October 26, 2010

Recent History of Drought vs Models

Yesterday, I talked about a recent review paper by Aiguo Dai which looked at projections for the Palmer Drought Severity Index (PDSI) in a variety of IPCC model runs for the twenty-first century.  The maps there are rather scary, and I think anyone who really gets what they say would very much prefer that the paper be wrong - preferably badly wrong.  There are of course multiple ways that the paper could be wrong, in principle, including at least:
  1. Dai could have some basic screw-up in his analysis such that other scientists would not produce the same results.
  2. The PDSI could be a bad indicator, particularly when applied to the output of climate models, such that the maps are actually quite misleading.
  3. The "average" of the climate models could be distorted by a few wacky models that produce nonsense (at least as input to the PDSI calculation).
  4. The state of climate modeling in general is quite imperfect, and perhaps the models simply cannot accurately forecast drought in particular - they all fail in the same way because climate science in general simply doesn't understand the climate well enough.
In general, it's not the case that all peer-reviewed papers are correct - scientists are human and make errors, both individually and (less often) collectively.  However, scientists are trained and expected to try to analyze all the ways their work might be wrong, and either show that they've taken care of the issue properly, or openly and honestly acknowledge that this wasn't possible at this time and further work of some kind is required to strengthen the evidence for the paper's hypothesis (there is always further work that could be done).  So the Dai paper has a variety of discussions and analyses that speak to points 1) to 4) above to varying degrees.  I've also read three other recent papers with related analyses that are worth discussing in detail (so far, hopefully I will shortly find others).

Today, I'd like to begin by looking specifically at one of the important sanity checks, which is the comparison to observation-based PDSI numbers.  Obviously, if one were to rely on model forecasts about the future, one would like to believe that the model could correctly reproduce the past and the present - at least roughly.  Dai himself is the keeper of a global dataset for historical PDSI numbers (any technically inclined reader can get it here if you want to play with it), and the paper contains several analyses of it.

The most important figure is this one (7c in the paper):

This shows the average trend in the author's preferred (more modern) version of the PDSI calculation between 1950 to 2008.  Recall from yesterday that in terms of the absolute level, 0 is normal, positive values are unusually wet, negative values are droughts, with -4 being a very severe drought and -6 being dustbowl-biblical (although clearly, it's got to depend to some degree on where you start from - a severe drought maybe worse if you started out pretty arid to begin with).  In the figure above, what we are looking at is the trend in the PDSI expressed as a change over fifty years.  So if you took all the PDSI figures for a particular spot, placed them on a time graph, fitted a straight line, and then expressed the slope of the line as so many PDSI units per 50 years.  Then greens and blues are places that have been getting less droughty, while yellow regions have gotten 0-1 points worse on the Palmer scale, pale orange 1-2 worse, and so on.

So what this says is that a lot of places have had a pretty material trend towards more drought since the 1950s - much of the Amazon, most of Africa, the Mediterranean, parts of Russia, eastern China, and eastern Australia have all been getting drier.  You will recognize how some of the news flow fits in these patterns - the Sahel drought of the late twentieth century, the recent record breaking Australian drought, fires in the Mediterrean. Notable exceptions are most of the United States (which seems cosmically unfair as the biggest emitters), Argentina, parts of the Sahara, Scandinavia.

So we can compare this trend figure to the IPCC-model-based figures.  Those are done by running the models through the twentieth century with the known forcings (carbon dioxide concentrations, changes in solar output, etc), and then running out through the twenty-first with the emissions scenario of choice.  Then Dai has taken model output and computed PDSIs.  For the same PDSI definition as just shown above, the model results for the twentieth century through the present are as follows.

Figure 10a) shows the situation in 1950-1959:

This is absolute PDSI levels on the following scale (which we saw yesterday):

So that was condition normal - the odd drought around the world, but nothing too disastrous.  By the 1975-1984 period, it should only have been a little worse according to the models (fig 10b):
While figure 10c) shows the situation for 2000-2009 (ie now).

So the models are saying that things should definitely have taken a turn for the worse by now.  The Mediterranean, much of China, the US, the Amazon should already be at Palmer -2 to -3, with spots of -4.  Let's compare that to the picture of the trend in the observations again.  These aren't exactly the same thing, since one is absolute PDSI, while the other is the change over 50 years, but given that we started not too far from zero in 1950, they are roughly comparable.  Here's the two together - observations on top, models on the bottom, each with their respective scales:

Note that these pictures were not designed ideally for this comparison, and in particular the color scales are not identical.  The stronger colors in the lower one are to some extent an artefact of the scale differences - with an extra kind of orange - so you'll have to try to make an allowance for that.

To my mind, the big picture conclusion is that the models are broadly getting it right that we have moved into a more drought-prone world, have the right general order-of-magnitude for the change, but appear to be making some significant errors about the regional distribution.  The US has been better off in reality than the models say, the Australian drying is not localized to the east in the models, reality has been harder on Siberia and Canada than the models said, etc.

So it's clear that the last generation of climate models are a distinctly imperfect guide to the future of drought, but at the same time, I don't think an open-minded observer can stare at the comparison and conclude from it that the models are worthless and their predictions should be entirely discounted.

In particular, the United States appears to be rather more blessed in reality than model-land.  Dai addresses this in a rather off-hand comment when he says:
Coupled climate models used in the IPCC AR4 project increased aridity in the 21st century, with a striking pattern that suggests continued drying over most of Africa, southern Europe and the Middle East, most of Americas (except Alaska, northern Canada, Uruguay, and northeastern Argentina), Australia, and Southeast Asia. Some of these regions, such as the United States, have fortunately avoided prolonged droughts during the last 50 years mainly due to decadal variations in ENSO and other climate modes, but people living in these regions may see a switch to persistent severe droughts in the next 20–50 years, depending on how ENSO and other natural variability modulate the GHG-induced drying.
So he's telling Americans that he thinks we just got lucky so far and he doesn't think our luck is necessarily going to last.  Clearly we need to take up this question of natural variability in the near future.

Note: This post is part of the Future of Drought Series on Early Warning.


Anonymous said...

Shouldn't it be quantifiable how much the models are missing it for any particular region & for the whole?

porsena said...

I don't think an open-minded observer can stare at the comparison and conclude from it that the models are worthless and their predictions should be entirely discounted.

On the other hand, I feel more comfortable than I did after yesterday's post. Maybe that's not a good thing, but thanks for today's analysis.

Stuart Staniford said...

Porsena - yeah, me too.

Emmaliu said...

I really enjoyed your review on Dai's paper. I also read his paper while I was doing drought research last year, but I didn't think as deep as you did back then. PDSI does have many disadvantages as you mentioned in the previous post, and these disadvantages might make it PDSI not a good indicator for drought prediction. I think I've read some papers talking about PDHI being a better real-time drought index.