Intro

(without broader/theoretical context…)

Mimetes fimbriifolius is a large charismatic shrub endemic to the Cape Peninsula. The species has thick protective bark and large individuals generally survive fire events, while seeds are dispersed by ants and persist in soil-stored seedbanks. Many individuals that survived the March 2015 Silvermine fire have since died and we would like to know if this mortality is related to microclimate as determined by topography (e.g. aspect, slope, elevation). We have detailed 30m gridded data layers for the area for topography, vegetation (from pre-fire satellite NDVI), and microclimate, based on spatial interpolation from 5 years of data from 100 stations. We propose sampling plots that span the range of topographic and microclimatic variation within Silvermine, recording the presence of M. fimbriifolius and scoring individuals as “Alive”, “Killed in the fire or died shortly after” (as evidenced by charring or brittle grey leaves), or “Died since the fire” (as evidenced by tan, flexible leaves).

Our questions (and proposed analyses):

  1. What is the species’ preferred habitat? - GLM with binomial distribution (i.e. logistic regression) of presence/absence data by plot/grid cell with GIS layers as predictors (topographic etc vars) - GLM with poisson regression of count data by plot/grid cell with GIS layers as predictors (topographic etc vars)

  2. Can we predict fire-induced mortality? - GLMM with binomial distribution for the survival/mortality of individuals (died post fire lumped into “survived”), with plot as a random effect, and with GIS layers and individual level variables (e.g. stem diameter, distance to nearest conspecific) as fixed effects

  3. Can we predict post-fire mortality and is it higher in “hotter/drier” topoclimates? - GLMM with binomial distribution for the survival/mortality of individuals (died in fire excluded), with plot as a random effect, and with GIS layers and individual level variables (e.g. stem diameter, distance to nearest conspecific) as fixed effects

  4. Is post-fire mortality within the 2015 burn scar higher than in unburnt areas? - This requires sampling unburnt areas. Data can be included in above analyses with burnt/unburnt as an additional factor in the model

  5. Are the constraints on the species’ occurrence the same as those that determine mortality during or after the fire? - Compare the outcomes of the models

We will need test the accuracy of using leaf conditions as an indicator of our mortality classes using long term data from Mimetes Valley. There are many simple ways of testing this.

We’re planning on stratifying the sampling using the existing Protea Atlas locations, so it should be easy to score the population size within a 500m radius as 0, 1-10, 11-100, 101-10000, 10000+, allowing us to do a very course test of changes in numbers/distribution over the past ~20 years.


Overview

Let’s have a look at the Protea Atlas distribution on the Cape Peninsula and within the March 2015 Silvermine fire scar.


## OGR data source with driver: ESRI Shapefile 
## Source: "/Users/jasper/Documents/GIS/CapePeninsula/TMNP/1962-2016 v1.1/TMNP_fires_1962_2016.shp", layer: "TMNP_fires_1962_2016"
## with 694 features
## It has 7 fields


There are 1609 M. fimbriifolius records across 5578 Protea Atlas Plots on the Peninsula, and 483 and 1265 within the March 2015 fire scar.



Environmental correlates?

Let’s look at the spatial data.



There’s strong collinearity among some of the covariates so let’s drop the worst ones (dem and TRI) and then see how well the Protea Atlas Plots sample the variation in covariates.



What if we fit GLMs (logistic regression) at the two scales?


#Set up GLM for species presence/absence
fit <- MCMClogit(Occurrence ~ janmax + TPI + slope + julrad + ndvi2014, data=datf, burnin = 5000, mcmc = 20000, thin=10)
summary(fit)
## 
## Iterations = 5001:24991
## Thinning interval = 10 
## Number of chains = 1 
## Sample size per chain = 2000 
## 
## 1. Empirical mean and standard deviation for each variable,
##    plus standard error of the mean:
## 
##                Mean      SD  Naive SE Time-series SE
## (Intercept) -4.9925 0.66978 0.0149766      0.0221353
## janmax       0.1685 0.02485 0.0005556      0.0008218
## TPI         -0.0332 0.06682 0.0014941      0.0021877
## slope       -0.6493 0.07470 0.0016705      0.0025635
## julrad      -0.2370 0.07777 0.0017389      0.0026105
## ndvi2014    -0.4024 0.07929 0.0017730      0.0026758
## 
## 2. Quantiles for each variable:
## 
##                2.5%     25%      50%       75%    97.5%
## (Intercept) -6.3320 -5.4189 -4.97733 -4.527782 -3.71938
## janmax       0.1206  0.1513  0.16800  0.184620  0.21900
## TPI         -0.1658 -0.0754 -0.03487  0.008617  0.09933
## slope       -0.7990 -0.6997 -0.64734 -0.596871 -0.50587
## julrad      -0.3911 -0.2889 -0.23858 -0.181410 -0.08585
## ndvi2014    -0.5567 -0.4558 -0.40155 -0.346797 -0.25444
plot(fit)


Where the posterior density shows little overlap of 0 indicates that the variable is important. So the model exploring the presence/absence of M. fimbriifolius reveals a weak negative effect of both mean maximum January temperature and the topographic position index (TPI - a measure of whether the focal cell sits higher or lower than its 24 surrounding cells (5x5 neighbourhood)), and significant negative effects of mean minimum July temperature, slope and maximum NDVI in the year preceeding the fire.


#Set up GLM for species abundance

fit <- MCMCpoisson(Abundance ~ janmax + TPI + slope + julrad + ndvi2014, data=datf, burnin = 5000, mcmc = 20000, thin=10)
summary(fit)
## 
## Iterations = 5001:24991
## Thinning interval = 10 
## Number of chains = 1 
## Sample size per chain = 2000 
## 
## 1. Empirical mean and standard deviation for each variable,
##    plus standard error of the mean:
## 
##                 Mean       SD  Naive SE Time-series SE
## (Intercept) -3.18706 0.171918 0.0038442      0.0056241
## janmax       0.18666 0.006277 0.0001404      0.0002057
## TPI         -0.07789 0.012664 0.0002832      0.0003861
## slope       -0.64464 0.015310 0.0003424      0.0004996
## julrad      -0.34782 0.016698 0.0003734      0.0005579
## ndvi2014    -0.21984 0.013565 0.0003033      0.0003988
## 
## 2. Quantiles for each variable:
## 
##                2.5%      25%      50%      75%    97.5%
## (Intercept) -3.5210 -3.29556 -3.19162 -3.07184 -2.84472
## janmax       0.1741  0.18251  0.18674  0.19067  0.19884
## TPI         -0.1025 -0.08644 -0.07803 -0.06907 -0.05374
## slope       -0.6738 -0.65522 -0.64487 -0.63445 -0.61307
## julrad      -0.3803 -0.35887 -0.34738 -0.33656 -0.31411
## ndvi2014    -0.2466 -0.22894 -0.21943 -0.21102 -0.19280
plot(fit)


The results looking at abundance are similar.