Constructs the train and test splits used to estimate predictive skill.
Usage
resample_scheme(
p,
method = c("spatial_block", "leave_location_out", "forward_season", "random"),
k = 5,
buffer = 0
)Arguments
- p
An
agri_project()with features built.- method
"spatial_block"groups units into compact blocks by k-means on their coordinates and holds out whole blocks;"leave_location_out"holds out whole units;"forward_season"trains on past seasons and tests on the next, never using the future to predict the past;"random"ignores structure entirely.- k
Number of folds.
- buffer
Exclude training rows whose unit lies within this distance of any test unit. Kilometres when the project is geographic, otherwise coordinate units. Buffering removes the residual optimism that blocking alone leaves at block edges.
Why the default is spatial
Agricultural observations near one another share weather, soil and management. Under random k-fold cross-validation a test point almost always has a near-duplicate in the training set, so the estimate answers "how well does this interpolate between my own plots" when the question asked is usually "how well does this predict somewhere new". The gap between the two is often large. Roberts et al. (2017) set out the problem and the blocking remedies; Meyer and Pebesma (2021) show how far a model can be trusted outside the space it was trained in.
"random" remains available, and train_model() reports both so the
difference can be quantified rather than argued about.
References
Roberts, D. R. et al. (2017) "Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure." Ecography 40, 913-929. doi:10.1111/ecog.02881
Meyer, H. & Pebesma, E. (2021) "Predicting into unknown space? Estimating the area of applicability of spatial prediction models." Methods in Ecology and Evolution 12, 1620-1633. doi:10.1111/2041-210X.13650
Examples
p <- agri_project(demo_agri_data(n_units = 12, n_seasons = 2))
p <- add_climate(p, source = "demo")
p <- phenology_windows(p)
p <- build_features(p)
resample_scheme(p, method = "spatial_block", k = 3)
#> <agri_resample>
#> method : spatial_block
#> folds : 3
#> test n : 8, 8, 8