rvanmazijk / cape-vs-swa

Open access repository for (some) data-sets, reproducible analyses, conference slides and manuscript for a publication based on my BSc Hons project (in press in Journal of Biogeography).

Ruan van Mazijk (he/him)

• Research & Publishing Intern @ C4 EcoSolutions 🌱🌍
• Ecologist, evolutionary biologist & botanist 🌿🌾
• BSc Hons, MSc (submitted) from the Department of Biological Sciences, University of Cape Town
• 1 page CV
• Email: [email protected]

Find me elsewhere:
ResearchGate | Google Scholar | ORCID
Twitter | LinkedIn

cape-vs-swa's Issues

Plot richness ~ enviro + roughness

(#15)

Run BRTs at 3QDS scale

Update paths in README

Re: R/02_analyses/ etc.

Update GitHub Pages site for this project

Species turnover note

(From Thu, 25 Jul)

This metric of turnover has the same problems as T_QDS / S_HDS and Jaccard distance:

HDS %>%
  mutate(euc_dist = sqrt(2*(add_turnover - mean_QDS_richness)^2)) %>%
  ggplot(aes(euc_dist, fill = region)) +
    geom_histogram(bins = 20, position = position_dodge())

Conclusion: scrap

Use jackknife aot bootstrap for CLES and U-tests

Following a chat with Tony earlier today, we think it is best that I rather use jackknife resampling, so as to really get at the issue of unequal number of pixels at different spatial scales when comparing Cape and SWA roughness.

What I was doing before:

bootstrap_sample <- function(x, n) {
  t(repeat(n, {
    sample(x, size = length(x), replace = TRUE)
  }))
}

Cape %<>% bootstrap_sample(n = 1000)
SWA %<>% bootstrap_sample(n = 1000)
CLES(SWA, Cape)

Rather, I should (i) limit resamples to the lowest common number of pixels (i.e. ca. 600 for 3QDS as the limiting size) and (ii) sample without replacement to make the jackknife distribution more "null". Also, Tony pointed out that I needn't resample my roughness-value distributions, but simply resample directly from the pairwise comparison matrix. This will save on computation time because I will only have to make that matrix once.

Thus:

jackknife_sample <- function(x) {
  sample(x, size = length(threeQDS(x)), replace = FALSE)
}
CLES_jackknife <- function(x, y, n) {
  pw <- matrix(nrow = length(rows), ncol = length(cols))
  rownames(pw) <- x
  colnames(pw) <- y
  pw_comparisons <- pw
  for (i in 1:nrow(pw) {
    for (j in 1:ncol(pw) {
      pw_comparisons[i, j] <- 
        rownames(pw)[[i]] < colnames(pw)[[j]]
    }
  }
  CLES_values <- vector(length = nrow(pw) * ncol(pw))
  for (n in 1:1000) {
    rows <- jackknife_sample(1:nrow(pw))
    cols <- jackknife_sample(1:ncol(pw))
    jackknifed_pw <- pw_comparisons[rows, cols]
    jackknifed_pw %<>% as.vector()
    CLES_values[[n]] <- sum(jackknifed_pw), na.rm = TRUE) / length(jackknifed_pw)
  }
  CLES_values
}

CLES_jackknife(SWA, Cape, n = 1000)

Change figure panel labels to lower case

Note that cellular roughness data for _that_ PCA incl. cells w/ < 4 sub-cells, but only those PC-values from cells w/ == 4 sub-cells analysed

Which "SoilsGrid250m" data?

Which "SoilsGrid250m" folder am I using for Cape vs SWA publication---the 4mb ones or the 300kb ones?

See RUAN_UCT/GIS/SoilGrids250m/*/depths-averaged/?/ for the smaller files currently in this project repo under data/derived-data/soils/.

Lozada-Gobilard S, Stang S, Pirhofer-Walzl K, Kalettka T, Heinken T, Schröder B, Eccard J, Joshi J (2018) Environmental filtering predicts plant-community trait distribution and diversity: Kettle holes as models of meta-community systems. Ecology and Evolution, in press.

Math-ify "QDS richness" etc. consistently in figure axes/labels

Should be S_QDS etc. instead

Test

Delete packrat/lib*/

This takes up Gbs of space!! (See c28b8f9)

Will make a branch just before 8ef150b to delete these directories.

Aspects to the study to maybe discuss

• Autocorrelative investigations are worthwhile
• Temporal range of your data sources << that of the question you're asking

AND:

• Cramer & Verboom 2016's methods § is NB!
• Linder paper in review by Tony is also NB!

Make better table and figure presentation when rendering ms to PDF

• Fix PNG figure sizes (currently over edge of page)
• Make prettier tables
  • bookdown has ugly default tables w/ kable::, so rather use kableExtra
• Fix LaTeX equation display in tables
• Fix ** use in data sources table---rather use kableExtra::group_rows()

Plot forms of BRT-learnt relationships between environment/heterogeneity & richness/turnover

Fix shapefile imports

Throws weird error/warning:

Warning message:
In readOGR(here::here("data/derived-data/borders/GCFR_QDS/")) :
  First layer lon read; multiple layers present in
/Users/ruanvanmazijk/projects/Cape-vs-SWA/data/derived-data/borders/GCFR_QDS, check layers with ogrListLayers()

Maybe is fine?

Hard reset to remove /old-images/ from repo

Not all are open source friendly!

Setup mirror of repo on FigShare

A "final product mirror" as opposed to the development repo that this is.

Maybe DataDryad?

Think also about UCT FigShare system (ZivaHub)

Manuscript note

• Tony is still working on the intro, but has put it aside for a bit as he is busy this week.
• When he gets back to it, he will chat to me about my methods section (and cutting it down)
• Then we will write the results.

Fix bookdown rendering

Getting this error on call bookdown::render_book("index.Rmd")

Error in x[i] <- sprintf("<a href=\"%s#%s\"", filenames[which.max(lines[lines <=  : 
  replacement has length zero

Neither editing the contents of index.Rmd to bare minimum, removing _output.yml, nor excluding all body .Rmds helps.

Save to bootstrap samples to disc for 0.05º roughness analysis

Because of the huge memory overhead when there are so many pixels at this resolution---causes hang & crash (see 8aacb71)

Cont. writing SIs

1. Collinearity among environmental predictor variables
2. Selecting representative BRT-models from sets of replicates
3. Testing for concordance in ranked lists

Outline talk

Print abstract & keywords into YAML header from file

In manuscript/index.Rmd, I currently have the (multiline!) abstract and keywords inside the abstract: argument of the YAML header:

abstract: |
  | **Aim** Foo
  | ** Location** Bar
  | **Taxon** Wizz
  |
  | _Keywords_
  | Foo, bar, wizz

This is pretty ugly.

I want to try and keep this stuff in, say, _abstract.Rmd, and dump it on render into the YAML header:

abstract: '`r readr::read_file("_abstract.Rmd")`'

Something like that would be great. I've tried it, but it creates an odd error wherein it is as if the whole YAML header is absent (thus giving no title, author, etc.), and throwing the error:

This document format requires a nonempty <title> element.
  Please specify either 'title' or 'pagetitle' in the metadata.
  Falling back to '_main.utf8'

Practice talk

Fix logo display in README.md on GitHub Pages site

![](logos/...) works well on GHP, as I see in my MSc's site, but then I lose size customisation (the { width=... } does not parse on GHP).

<img src="logos/..." width=...> and ![](logos/...) work for the README on the repo display, but <img ...> breaks for GHP.

Perhaps manually reduce size of huge logos and then simply use ![](...)?

Cont. manuscript in MS Word w/ Mike & Tony

BW friendly Cape (orange) and SWA (blue) colours

This is problematic. Compare colour

to black and white

for the roughness distribution panels.

I imagine this is also a problem for all of Figure 2 and 3 too.

Add geographically weighted regressions

In #11 mentioned the importance of autocorrelative structure in space.

Want to apply GWR for HDS richness ~ avg QDS richness & turnover models, and also (maybe) for richness/turnover ~ environment/heterogeneity models (eventually!).

Mention lack of need for autocorrelative investigations

Due to autocorrelated predictors

Add maps from SAAB-AMA-SASSB-2019 talk to ms

Note, figure/table/equation in-text styling is also customisable

Using _bookdown.yml, you can specify how these things appear in text before a bookdown \@ref()-call. See https://bookdown.org/yihui/bookdown/internationalization.html for more (copied YAML below.

language:
  label:
    fig: 'Figure '
    tab: 'Table '
    eq: 'Equation '

Or rather, if you wish:

language:
  label:
    fig: 'Fig. '
    tab: 'Table '
    eq: 'Eqn. '

Re-clone Cape-vs-SWA.git -> MacBook Air

Add mixed (random) effects models

This would be especially useful for the HDS richness ~ avg QDS richness & turnover models (and maybe even the richness-environment models---unless we use BRTs).

This is is because traditional multiple linear regressions (that I have used for the richness-turnover stuff so far) assumes independence between observations. Mixed effects models structure error by the categories, not having this assumption---seemingly ideal for my comparing two regions.

Testing

Get up-to-date figures from ms, SI -> slides

• Context map (#122)
• Figure S4: PCA biplots (modified version from MEDECOS 2020) (#123)
• Figure 1a–d: richness, PC1 maps (#123)
• Figure 4: richness vs PC1 (#124)
• Figure S5: PC1-based outliers (#124)

Separate repos into analysis, manuscript, slides

Improve richness ~ enviro + roughness **more**

(#15, #43)

What I have done so far:

Richness @QDS ~ Enviro @QDS + Roughness BetwQDS

What I think would be better:

Richness  @QDS ~ Enviro  @QDS + Roughness  WithinQDS([email protected])
Richness  @HDS ~ Enviro  @HDS + Roughness  WithinHDS([email protected])
Richness @3QDS ~ Enviro @3QDS + Roughness Within3QDS([email protected])

And

Turnover  @HDS ~ Enviro  @HDS + Roughness  WithinHDS([email protected])
Turnover @3QDS ~ Enviro @3QDS + Roughness Within3QDS([email protected])

Diversity index as measure of heterogeneity in discrete soil class layer?

• Search for SWA soil classification layer
• Get ARC LandTypes layer for GCFR from Mike

Mention conservation importance

A variety of habitats obviously NB to species richness conservation

"With & without the Kogelberg" analyses no longer needed

Running these models:

foo1 <- lm(HDS_richness ~ PC1, HDS)
foo2 <- lm(HDS_richness ~ PC1, HDS[HDS$HDS_richness < 2500, ])
plot(HDS_richness ~ PC1, HDS, col = factor(HDS$region))
abline(foo1)
abline(foo2, lty = "dashed")

foo1 <- lm(QDS_richness ~ PC1, QDS[QDS$QDS_richness < 2000, ])
foo2 <- lm(QDS_richness ~ PC1, QDS)
plot(QDS_richness ~ PC1, QDS, col = factor(QDS$region))
abline(foo1)
abline(foo2, lty = "dashed")

Yields slopes from foo1 and foo2 that are almost identical.

Conclusion: the "outliers" of the Kogelberg (or the bits of the QDS raster that kind of touch the Kogelberg, because a lot of the "real" Kogelberg is excluded from the final dataset due to the edge buffer) are not a big deal!!!

Remake figures to be more talk-friendly, as needed

Re: #115.

• Figure 1a–d: richness, PC1 maps (#123)
  • Vertical colour ramps
  • Coloured region headings
• & all figures w/ new colour scheme (#127)? Namely:
  • Context map (#122)
  • Figure S4: PCA biplots (#123)
  • Figure 4: richness vs PC1 (#124)

Collate "core" code for analyses & maybe make this repo a "dev" repo?

Repeat HDS richness ~ mean QDS richness and turnover w/ GWR

(#15)

Update slides' title to match abstract submitted/final paper name?

Re: #120.

Or at least include the citation-proper on the title slide?

Testing

Re-install MS Word -> MacBook Air

Make "adverstisement" page for this project on my GHP blog, or as own repo

E.g. index, preliminary-analyses, slides, proposal

Basically a description, some results, figures, link to the RG project

Change figure panel labels to lower case

Wiley Author Guidelines specify that panels should be numbered "(a) (b) (c)".