integrate drought

[teixeirak]

The main scientific element that coauthors felt was missing is inclusion of drought.

From Loïc:

see issue #61

From Justin:

My big picture feedback is similar to Loic's in that it is drought focused. I think there are two drought components missing, one is the summer drought which it seems you all are thinking through. To add to this discussion, I will add that radial growth in the eastern US show pervasive drought legacies (paper attached) in growth mainly due to late season drought. So it may not only be that spring temp and summer temp are linked during the current year but also that autumn drought could have an impact the following year and may weaken the response to spring temperature, if that makes sense. Interestingly, other phenology responses at least at MMSF do not seem to have these legacies so this could be one possible contribution to the disconnect (2nd paper attached). The 2nd point that came to mind was when talking about warming springs in the future. At some point those warmer temps will increase water demand so one would expect an offset to the benefit of warmer spring temps at some point, right?

Neil also mentioned need to consider drought.

my follow-up question:

Do we think it’s sufficient to discuss drought, or do we feel that additional analysis is needed? Cameron did try a drought analysis on the dendro band data at one point (https://github.com/EcoClimLab/growth_phenology/issues/27), but we didn’t see much of a signal, and we dropped the idea. With the limited number of years with dendro band data, I don’t see how we could include drought in addition to spring temperatures without over-fitting, but I’d love to hear any clever ideas for bringing drought into that analysis. We could of course test for effects of both in the tree ring analysis, and that might be worth doing.

From my work on another paper (ForestGEO tree-rings analysis, in which many of you are involved), I’m thinking that it may make sense to consider two elements of drought: soil moisture (operating over relatively long seasonal windows and shaped by both precip and PET) and VPD (mainly a temperature effect operating over shorter seasonal windows; see the attached). This could either simplify (if we focus on just the latter) or complicate the analysis/ message here.

[camerondow35]

(Sorry its taken so long for me to get to this! It sure doesn't feel like its been 26 days.)

How about we incorporate drought into our mixed model? I'm thinking a fixed variable for previous year autumn drought and another for current year summer drought. I've done a prelim version of this using raw precip values and only 5000 iterations instead of our 10000 (my laptop cant handle 10k). Neither the autumn or summer precip variables were significant over our study years. I'm guessing we should use a different variable than precip sum, but the code is set up to plug variables in pretty easily.

If the drought effect comes out non-significant over our study years, can we report it wasn't a factor in our results, then use that to discuss the potential effects over a longer period? And If it comes out significant than we can report that! I'm wondering if there will be alot of disagreement about the correct variable to use if we do this though...

[teixeirak]

Just looking back at old emails to first answer the question of whether we need an analysis.

Do we think it’s sufficient to discuss drought, or do we feel that additional analysis is needed?

Justin thinks is sufficient to just discuss it. I'm not sure anyone else responded on that...

[teixeirak]

If we do analyze drought, I'd go with something linked to the temperature side of drought (Tmax or PET), or something that integrates both (SPEI). Given that we show this strong negative effect of Tmax, it seems most parsimonious to analyze Tmax. I'm going to try to do a bit of reading and provide some better comments here.

[teixeirak]

@camerondow35 , I'm still not sure exactly what we should do here, but some thoughts:

1- It would be good to analyze the correlation between spring T and summer T and SPEI, both for the dendro band years (higher priority) and for the tree-ring analysis years. This is a first step at describing whether warmer springs are in fact associated with more summer drought in our data set.

2- For the dendro bands, I don't know that we'd want to add drought to the analysis because (1) we have just enough years of data to support a regression against one climate variable (spring T). I worry about over-fitting if we try to include two; and, more fundamentally, (2) many of the variables apply to growth before the main months of summer drought. DOY75 occurs by early July, so July and August climate are relevant only to ∆DBH. Of course, we have much better annual growth data from tree rings (see next point). May-June Tmax or SPEI (especially June) would be the relevant climate metric to consider, although those shouldn't be included for DOY25 or some DOY50. But we could test whether DOY75 or L_PGS are responsive to drought.

3- For the cores, we could reasonably analyze the joint effects of spring T and summer drought. Is this analysis something you'd be up for doing? I'd envision a mixed effects model with fixed effects of April Tmax and a drought metric ( May-Aug Tmax or SPEI) and a random effect of chronology. If you want to do this, let's open a new issue and work out the details.

[camerondow35]

3- For the cores, we could reasonably analyze the joint effects of spring T and summer drought. Is this analysis something you'd be up for doing? I'd envision a mixed effects model with fixed effects of April Tmax and a drought metric ( May-Aug Tmax or SPEI) and a random effect of chronology. If you want to do this, let's open a new issue and work out the details.

After reading Buermann I think we should definitely do some analysis here. It looks like eastern USA is predicted to have positive correlations between spring temps and summer/autumn productivity. I really don't like using satellite data to predict productivity though. It seems too difficult to discern between shrubs and other annual plants under the canopy to make any conclusions on tree productivity - which is my guess for what's happening here. If we do an analysis and discuss Buermann's results I think it'll bring home the message that this extra carbon isn't being stored in wood.

[teixeirak]

Ok, I broke #3 off as a separate issue (#66) and will address it there, as its big enough to be stand-alone.

In this issue, I think item 1 in the list above is still worth addressing.

[camerondow35]

1- It would be good to analyze the correlation between spring T and summer T and SPEI, both for the dendro band years (higher priority) and for the tree-ring analysis years. This is a first step at describing whether warmer springs are in fact associated with more summer drought in our data set.

I'm not very familiar with SPEI. Would I use the 1-month or some other month value here?

[teixeirak]

There's no "right" answer as to what time frame to use with SPEI, but 1 mo is generally not going to be the best for capturing the time scale affecting tree growth. As in issue #66 , I'd recommend trying at least a couple different scales: maybe 4-mo and 12-mo ending in August.

[camerondow35]

Ok, Bianca explained how to use the SPEI dataframes to me but it's been a while. What you're suggesting is to use the August 4-month and 12-month values in this dataframe, and comparing them to spring T (April or March-May) and summer T (June, June-August)? So if I were to plot them against each other it would look like this?

[teixeirak]

Yes, something like that. Regression lines should just be linear, though. There aren't enough data to support a spline fit.

[teixeirak]

It looks like April Tmax and SPEI are pretty independent for our dendrobands data at SCBI.

[camerondow35]

Yeah, that would agree with my previous attempts of assessing the drought stress at SCBI in these years. FYI none of the April linear models are sig.

May and June both seem important in the SCBI quilt plots so I plotted all 3. Is this what you were hoping for? I can't say I 100% understand these!

[teixeirak]

FYI none of the April linear models are sig.

Okay, that's the one that matters. There are non-sig trends in the direction that would point to drought for 4- and 6-mo SPEI, though. What's the p-value on those?

[camerondow35]

p = 0.35 and p = 0.37 for 4 and 6-month respectively

[teixeirak]

ok, thanks... so not close.

[teixeirak]

I wonder if we should check May - June - July SPEI too, though. Those are the most critical months.

I'm not sure exactly how SPEI is calculated. All of the variables you considered would include conditions in those months, but the SPEI values for the months in which growth occurs may be more relevant.

[teixeirak]

From this website,

"With a value for PET, the difference between the precipitation (P) and PET for the month i is calculated: Equation 7.,

which provides a simple measure of the water surplus or deficit for the analyzed month.

The calculated Di values are aggregated at different time scales, following the same procedure as for the SPI. "

It does not explain how SPI is calculated.

[teixeirak]

This pub explains how SPI is calculated.

[camerondow35]

Are SPI and SPEI the same thing?

Maybe the calculation is somewhere in Bianca's script? I can't see it at first glance though....

[teixeirak]

From McKee et al:

The Standardized Precipitation Index (SPI) iscalculated in the following sequence. A monthlyprecipitation data set is prepared for a period of mmonths, ideally a continuous period of at least 30years. A set of averaging periods are selected todetermine a set of time scales of period j monthswhere j is 3, 6, 12, 24, or 48 months. These representarbitrary but typical time scales for precipitation deficitsto affect the five types of usable water sources. Thedata set is moving in the sense that each month a newvalue is determined from the previous i months. Eachof the data sets are fitted to the Gamma function todefine the relationship of probability to precipitation.Once the relationship of probability to precipitation isestablished from the historic records, the probability ofany observed precipitation data point is calculated andused along with an estimate of the inverse normal tocalculate the precipitation deviation for a normallydistributed probability density with a mean of zero andstandard deviation of unity

[teixeirak]

Are SPI and SPEI the same thing?

No (SPI doesn't account for evaporation), but they're aggregated across time scales in the same way.

[teixeirak]

Maybe the calculation is somewhere in Bianca's script? I can't see it at first glance though....

It would be in these lines. I'm not pro at R, but it looks like all months are weighted equally, which is my main question.

So, August 4-mo SPEI would be average Deficit (= PRE - PET) for May-June-July-Aug.

On the surface, that's probably what we want. But if you have a very dry May-June (when most growth occurs) and then a very wet July-Aug, SPEI would be normal but growth would be drought- affected. So it might make sense to try June SPEI as well, just to really convince ourselves there's no correlation.

[teixeirak]

Note that Aug 4-mo SPEI is completely independent of April T, whereas June SPEI would include Mar-April-May-June, thus capturing potential water deficit buildup from warm springs.

[camerondow35]

One column for each month SPEI (June, July, August SPEI). Legend text is slightly cut off but blue line is still 12-month SPEI value. P-value for June SPEI vs April Tmax are 0.38, 0.25, and 0.71. P-value for July SPEI vs April Tmax are 0.31, 0.21, and 0.77. So still negative trending but non-sig

[teixeirak]

Ok, thanks. You can drop the comparisons with May and June Tmax, by the way.

I'm now satisfied that we can say there's no correlations between April T and SPEI for the dendro bands time period at SCBI.

[camerondow35]

Ok great. My time with WiFi is nearing an end for today but I'll do HF and the tree cores (if I can) when possible.