poisotlab / biodiversityobservationnetworks.jl Goto Github PK

Currently blocked pending release, will do asap

In response to

Looks good, should we merge and then apply blue style w/ JuliaFormatter?

Originally posted by @gottacatchenall in #5 (review)

Definitely -- it might be a good idea to add a formatter file at the repo's root. I don't like everything in BlueStyle, but we do want a formatter.

Alternatively, can it be running at github actions time?

During the spatial metaweb working group meetings, we discussed a roadmap for the package development and writing of the package manuscript. This is what we planned (in chronological order):

• [ ] API simulation to batching
• v0.1 goes live
• Vignette and documentation
• Test if having species distributions in estimate is better (since it is already a part of the uncertainty score)
• GEO BON flex
• Cube algorithm
• Trade-offs in temporal sampling (costs)
• Input layer of the number of unknown interactions (and type ? score ?)
• Preprint (December 2022) - Methods (Pkg), Ecography (maps)
• Modify simulation so optimization

We should check off the ones that have already been done, and write the others as individual issues and milestones, when appropriate.

In order to write things like X |> seed(BalancedAcceptance()) |> refine(AdaptiveSpatial()), all seed and seed! methods need to return a tuple matching the arguments to the curried methods for refine and refine!.

This sounds like word salad but -- I'll make this happen when #5 is merged

• Error handling

  • Each invalid call should throw an error of abstract type BONException
  • Have a documentation section for each subtype of BONException to enable user debugging

• Algorithms

  • cube sampling (refiner)
  • Environmental uniquness via KNN (seeder)
    • Take the vector of layers at each cell in the raster and applies knn to each cell in the raster. choose the closest cell in the raster to the k cluster centers
  • fractal triad, or some scale dependence thing, maybe

• Documentation

  • A tutorial section that gives background info not specific to the package for new users
  • Uniqueness vignette includes KNNUniqueness and has more text
  • Vignette on cube smapling

bbox = (left=-83.0, bottom=46.4, right=-55.2, top=63.7);
h = convert(Float32, SimpleSDMPredictor(WorldClim, BioClim, 7; bbox...))
h |> seed(BalancedAcceptance)

fails with

ERROR: MethodError: no method matching seed!(::Vector{CartesianIndex}, ::BalancedAcceptance{Int64, Float64}, ::SimpleSDMResponse{Float32})
Closest candidates are:
  seed!(::Vector{CartesianIndex}, ::ST) where ST<:BONSeeder at ~/.julia/packages/BiodiversityObservationNetworks/IR0dX/src/seed.jl:30
  seed!(::Vector{CartesianIndex}, ::ST, ::Matrix{T}) where {ST<:BONSeeder, T<:AbstractFloat} at ~/.julia/packages/BiodiversityObservationNetworks/IR0dX/src/seed.jl:9

Apparently computers hate me and don't like sending transparent PNG

Example data for the vingette is up on Zenodo (10.5281/zenodo.7023280). Currently includes ESA land cover data for 2020, Global Human Modification of Terrestrial Systems dataset , and annual temp and precip from Chelsa Clim.

Hear me out: (BalancedAcceptance(numpoints=100)∘AdaptiveSpatial(numpoints=10))(X)

This specifies a complete pipeline for a matrix of entropy, and it's probably the most concise notation we can think of.

Little heads up that I've moved this package under the lab org so it reflects the contributors -- nothing else changes (tagging @gottacatchenall )

This looks great, but I came here wondering what BONs are!

What do you think of just using BiodiversityObservationNetworks.jl?

Julia package name guidlines:
https://pkgdocs.julialang.org/v1/creating-packages/index.html#Package-naming-guidelines

@gottacatchenall what is your input on the following additional way to use the package: rand(method, landscape, coordinates)?

Hear me out: what if we want to get 1000 points using a first method, and then 100 using a second method. Something like:

landscape |> rand(method1) |> rand(method2)

would be very cool to write, right? As far as I can tell the only change required would be to make sure that the methods return (landscape, coordinates), and maybe an additional method so that rand(method) is a function that unpack its arguments?

This would deviate from what rand does, so we may want to think about a rename? monitoringsites?

This package implements methods for optimizing the location of spatial sampling for biodiversity monitoring as part of the block 3 for the GEOBON manuscript project.

The following methods are primarily what I've seen in the literature for optimizing spatial sampling.

Opening this up as a place for discussion to decide on a final list of methods before implementing.

1. Fractal Triads Simpson et al 2021
Proposed to measure phenomena across scales. Of those listed here, this is the simplest, as it always recommends a set of nested triangles, and does not use any information about the landscape/species (e.g. an SDM) to design points.

2. Generalised random-tessellation stratification (GRTS) Stevens and Olsen (2004)
Argues spatially balanced sampling is best, where spatially balanced is defined as meaning both points are not clumped together, but also that all spatially contiguous subsamples of this total sample are also spatially balanced. Proposes a method to generate balanced samples better than random sampling.

3. Pivotal Method (Grafstrom 2011)
Proposed improvement to GRTS

4. Balanced acceptance sampling (van Dam-Bates et al 2018)
Proposed improvement to GRTS and Pivotal method

Methods below this line are different in that they use some spatial information to design samples (e.g. environmental predictors for an SDM, or an SDM with uncertainties)

5. Spatial Simulated Annealing van Groenigen 1998, and computational improvement by (Chen et al 2013). Optimizes spatial locations such that they minimize covariance among a set of spatial covariates (e.g. environmental raster data).

6 Adaptive sampling (Andrade-Pacheco et al 2020)
Prioritizes spatial location based on current uncertainty of measurement there

Needs an implementation of rand or _generate that handles SDMLayers as inputs, in particular, handling of nothing values

Whitening + PCA(?)