Shift and rotate

In this tutorial, we will apply the Shift and Rotate technique to generate realistic maps of predictors to provide a null sample for the performance of an SDM. We will also look at the quantile mapping function, which ensures that the null sample has the exact same distribution of values compared to the original data.

using SpeciesDistributionToolkit
const SDT = SpeciesDistributionToolkit
using CairoMakie

We will get some occurrences from the butterfly Aglais caschmirensis in Pakistan:

pol = getpolygon(PolygonData(NaturalEarth, Countries))["Pakistan"]
presences = GBIF.download("10.15468/dl.emv5tj");
extent = SDT.boundingbox(pol)

(left = 60.84437942504883, right = 77.0489730834961, bottom = 23.69452476501465, top = 37.05448532104492)

And then get some bioclimatic variables:

provider = RasterData(CHELSA2, BioClim)
L = SDMLayer{Float32}[
    SDMLayer(
        provider;
        layer = x,
        extent...,
    ) for x in 1:19
];
mask!(L, pol)

19-element Vector{SDMLayer{Float32}}:
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)
 🗺️  A 1604 × 1945 layer (1176577 Float32 cells)

As for the previous tutorials, we will generate some background points:

presencelayer = mask(first(L), Occurrences(mask(presences, pol)))
background = pseudoabsencemask(DistanceToEvent, presencelayer)
bgpoints = backgroundpoints(nodata(background, d -> d < 20), 2sum(presencelayer))

🗺️  A 1604 × 1945 layer with 1157282 Bool cells
   Projection: +proj=longlat +datum=WGS84 +no_defs

And finally train an ensemble model:

sdm = Bagging(SDM(RawData, DecisionTree, L, presencelayer, bgpoints), 50)
bagfeatures!(sdm)

{RawData → DecisionTree → P(x) ≥ 0.5} × 50

Let's look at the first prediction of this model:

train!(sdm)

heatmap(predict(sdm, L))
lines!(pol)
current_figure()