Nested cross-validation with spatial block CV and cluster CV

[gregor-fausto]

I'm setting up a nested cross-validation using spatial block cross-validation for outer folds and cluster cross-validation for inner folds. Interaction between these two types of cross-validations gives me the following error:

library(spatialsample)
library(rsample)
folds <- spatialsample::spatial_block_cv(boston_canopy, v = 3)
rsample::nested_cv(boston_canopy, outside = folds, inside = rsample::clustering_cv(v = 3, repeats = 1, vars = c('mean_temp','mean_heat_index_morning')))
#> Error in `map()`:
#> ℹ In index: 1.
#> Caused by error in `dist()`:
#> ! 'list' object cannot be coerced to type 'double'

The behavior that I'm trying to achieve is shown here:

library(spatialsample)
library(rsample)
library(sf)
#> Linking to GEOS 3.11.0, GDAL 3.5.3, PROJ 9.1.0; sf_use_s2() is TRUE
folds <- spatialsample::spatial_block_cv(boston_canopy, v = 3)
tmp<-list()
for(j in 1:nrow(folds)){
  x<-rsample::analysis(folds$splits[[j]]) %>% sf::st_drop_geometry()
  tmp[[j]] <- rsample::clustering_cv(data=x, v = 3, repeats = 1, vars = c('mean_temp','mean_heat_index_morning'))
}
folds$inner_resamples <- tmp
folds
#> #  3-fold spatial block cross-validation 
#> # A tibble: 3 × 3
#>   splits            id    inner_resamples   
#>   <list>            <chr> <list>            
#> 1 <split [458/224]> Fold1 <clt_fold [3 × 2]>
#> 2 <split [469/213]> Fold2 <clt_fold [3 × 2]>
#> 3 <split [437/245]> Fold3 <clt_fold [3 × 2]>

This example uses data from the spatialsample package. I've opened the issue in rsample because the error message seems to be related to how rsample::nested_cv interacts with the output of spatial_block_cv.

[hfrick]

Thanks for reporting this 🙌 spatialsample requires data to be an sf object. That has a "sticky" geometry column that other parts of tidymodels aren't necessarily expecting, see also tidymodels/spatialsample#159

At the moment, I'm unsure where it would be best to catch that. Possibly in spatialsample itself, so that rsample could avoid a dependency on sf.

For now, it might be best to tweak the data in each rsplit manually, if you are sure you want to drop the geometry column.

library(spatialsample)
library(rsample)

folds <- spatialsample::spatial_block_cv(boston_canopy, v = 3)

folds$splits <- purrr::map(folds$splits, function(x) {
  x$data <- sf::st_drop_geometry(x$data)
  x
})

nested_cv(outside = folds, inside = clustering_cv(v = 3, repeats = 1, vars = c('mean_temp','mean_heat_index_morning')))
#> # Nested resampling:
#> #  outer: `folds`
#> #  inner: 3-cluster cross-validation
#> # A tibble: 3 × 3
#>   splits            id    inner_resamples   
#>   <list>            <chr> <list>            
#> 1 <split [446/236]> Fold1 <clt_fold [3 × 2]>
#> 2 <split [422/260]> Fold2 <clt_fold [3 × 2]>
#> 3 <split [496/186]> Fold3 <clt_fold [3 × 2]>

^{Created on 2026-01-29 with reprex v2.1.1}