A good measure of drought severity is the Palmer Drought Severity Index, also called simply “PDSI.” One thing to keep in mind is that negative numbers indicated drought, the more negative the more severe, while positive numbers indicate wet conditions. Values of -2 or lower indicate moderate drought, -3 and below are severe drought, and -4 or lower marks extreme drought.
NOAA reports the PDSI for each month, for each of the 344 climate divisions covering the conterminous USA (the “lower 48 states”). I decided to apply PCA (principal component analysis) to look for patterns through time that are common to larger regions. I also decided to do so using smoothed data, in order to focus on longer-term trend patterns rather than short-term fluctuation patters, as I did here for temperature data.
Without further ado, here’s the 1st PC (principal component) for PDSI (Palmer Drought Severity Index):
The graph on the right shows the pattern over time, which is the most prominent one since this is PC1. The map on the left shows how strongly each climate division’s individual (smoothed) data matches that pattern; red means a match (correlation), blue means anti-match (anticorrelation), and larger dots mean stronger correlation or anticorrelation.
The most obvious feature of the time series is the prominent dip in the 1930s which, dipping negative, means enhanced drought conditions. That’s the drought which created what came to be called the 1930s “dust bowl,” which wreaked havoc on agriculture in much of middle America. The large red dots on the map show that the dust bowl drought was focused on a swath of middle America, mainly from the Dakotas down through Kansas and Oklahoma. If you want to know how this affected the lives of Americans in those areas at that time, read the book or watch the movie The Grapes of Wrath. It’s not a happy story.
The 2nd PC shows a different pattern:
The main feature of this pattern is an overall upward trend since 1895 (when the records begin). Red dots dominate the northeast, showing that conditions got wetter so drought has, overall, declined there in the past century and a half. Blue dots, signalling anticorrelation, dominate the southwest, showing that drought has increased there during the same time period. The general pattern agrees with what is expected from global warming, that wetter areas tend to get even wetter while drier regions tend to get more dry.
Here’s the 3rd PC:
The pattern is one of more drought very early, and enhanced drought in the 1950s-1960s but suppressed drought in the 1970s-1980s. This pattern shows strongest in the south, as far east as Alabama, while the opposite patterns shows up in a small region of the north mid-country centered on North Dakota.
The 4th PC shows enhanced drought during the 1920s:
It’s concentrated in western states, but again there’s a small region correlating with the opposite pattern, concentrated in Wyoming.
The main conclusion from all this is the overall trend pattern shown by PC2, with less drought in the northeast and more in the southwest. That pattern is further emphasized by the other PCs, especially the increase in southwestern drought. Increasing drought conditions are probably a major factor in the increase of wildfire damage in the western USA, and all changes to the water cycle tend to spell trouble for agriculture. California has been hit particularly hard, since it’s the state with the largest agricultural output.
This blog is made possible by readers like you; join others by donating at My Wee Dragon.
Open Mind 
Very interesting analysis. Just wondering how much of the data variability is explained by the first (and subsequent) principal components? Do the 3rd and 4th PCs add that much more information?
[Response: The first PC accounts for 30% of the variance, the 2nd for 18%, the 3rd 9%, the 4th 7%. Note that’s the variance of the *smoothed* data — there a lot of short-term variance which doesn’t even enter into the analysis.]
I’m surprised that you need 3 principal components to account more than 50% of the variance. Perhaps I’m surprised because I thought that autocorrelated data will tend to have more of their variance contained in the first couple of PCs, and I thought these data would be quite autocorrelated.
Perhaps also indicative of my ignorance: I thought having more variance accounted for by multiple PCs was cool because “there are more relationships and patterns to learn about”; but I think I read that seeing most of the variance encapsulated in the first PC was better because that confers more confidence regarding how the variables relate.
Don’t know if it’s the case here, but sometimes signals with periodic components need doublets of principal components (e-vectors or singular vectors, similarly) to explain their behavior or, equivalently, to reconstruct the signal well.
I got into a short dustup once with a climatologist/atmospheric scientist who, with a co-author in a paper claimed there was nothing extraordinary about recent California droughts because the Palmer Index was a broken one and when properly interpreted the ordinariness of California precipitation was clear. I fought back for a bit but then withdrew since I am no expert at this. I did resign my AGU membership over it and haven’t missed it, doing that since financial contributions to climate-related scholars is the most potent thing I can do. (And, oh yes, the saved monies were reallocated, this time to WHOI.)
But I wonder what those weaknesses in the Palmer Index are about, and if they are real, or if the Palmer Index is still pretty good despite blemishes.
Some basics here (apologies if you’re already familiar with such):
https://en.wikipedia.org/wiki/Palmer_drought_index
But note that there are actually–and contrary to the above article–several ‘Palmer Indices’, although the PDSI is by far the most commonly used:
https://www.ncdc.noaa.gov/temp-and-precip/drought/historical-palmers/
The issues around drought metrics are as I understand it still quite fraught, which is one reason that AR5 ended up walking back claims of increasing global drought made in AR4. I’m not sure what the current state of the art is in this regard, but I’d bet my bottom dollar there is some hard work going on to clarify things.
Thanks, @Doc Snow!
I was reasonably familiar with those.
Not wanting to put words in anyone’s mouth, but, in the spirit of what @WebHubTelescope often suggests, it seems to me there are a few, new, robust techniques for determining which parameters give the best predictions of these, with L{sub}2 boosting being one, and, my preference, Bayesian additive regression trees being another. Don’t need to concoct ad hoc indices any longer!