Alternative Site Selection

Introduction

Planned sampling locations sometimes turn out to be inaccessible in the field — private land, dangerous terrain, or logistical constraints. Rather than dropping those locations entirely, ss_alt_sites() finds environmentally similar replacement sites elsewhere in the study area, so the overall design keeps covering the same range of conditions.

The approach:

Generate a pool of candidate sites across the study area (ss_alt_candidates())
Optionally exclude candidates too close to the inaccessible sites, so replacements are spatially distinct (ss_alt_filter_buffer())
Compute environmental similarity between each inaccessible site and every candidate (ss_alt_similarity())
Rank and select the most similar candidates as alternatives (ss_alt_rank())

Theoretical Background

Similarity Metrics

ss_alt_similarity() supports three distance metrics between a target site and a candidate, converted to a [0, 1] similarity score (1 = identical):

Mahalanobis (default): accounts for correlation between covariates via the population covariance matrix. Most robust when covariates are correlated, but needs enough candidates to estimate a stable covariance matrix.
Euclidean: simple geometric distance after range-normalization. Faster, ignores covariate correlation.
Gower: handles a mix of continuous and categorical covariates via per-variable normalized differences.

Distance Buffering

ss_alt_filter_buffer() removes any candidate within min_distance of any inaccessible site, before similarity is computed. This avoids picking “alternatives” that are really just adjacent to the original problem location (e.g. the same inaccessible area).

Basic Usage

library(soilsampling)
library(terra)

set.seed(11)

nr <- 40
nc <- 40

r <- rast(nrows = nr, ncols = nc, xmin = 0, xmax = 1000, ymin = 0, ymax = 1000, nlyrs = 3)
names(r) <- c("dem", "slope", "ndvi")

xy <- xyFromCell(r, seq_len(ncell(r)))
values(r)[, "dem"] <- 100 + 20 * sin(xy[, 1] / 150) + rnorm(ncell(r), 0, 3)
#> Warning: [readValues] raster has no values
values(r)[, "slope"] <- abs(cos(xy[, 2] / 120) * 8 + rnorm(ncell(r), 0, 1))
values(r)[, "ndvi"] <- pmin(pmax(0.5 + 0.3 * cos(xy[, 1] / 200 + xy[, 2] / 250) +
  rnorm(ncell(r), 0, 0.05), -1), 1)

# A few "inaccessible" sites needing alternatives
inaccessible <- data.frame(
  x = c(200, 500, 800),
  y = c(200, 500, 800)
)
inaccessible
#>     x   y
#> 1 200 200
#> 2 500 500
#> 3 800 800

Running the End-to-End Workflow

result <- ss_alt_sites(
  r,
  inaccessible,
  method = "mahalanobis",
  n_alternatives = 3,
  min_distance_buffer = 100,
  seed = 123
)

result$alternatives[, c("site_id", "target_site_id", "similarity_score", "similarity_rank")]
#>         site_id target_site_id similarity_score similarity_rank
#> 1  candidate_46       target_1        0.9392170               1
#> 2 candidate_395       target_1        0.9289847               2
#> 3 candidate_858       target_1        0.9289847               3
#> 4  candidate_11       target_2        0.9575523               1
#> 5 candidate_953       target_2        0.9575523               2
#> 6 candidate_956       target_2        0.9575523               3
#> 7 candidate_119       target_3        0.9479020               1
#> 8 candidate_185       target_3        0.9479020               2
#> 9 candidate_477       target_3        0.9257395               3

Each row of result$alternatives is one candidate selected for one target site, ranked by similarity (similarity_rank = 1 is the best match).

Function Walkthrough

Generating Candidates

candidates <- ss_alt_candidates(r, n_candidates = 300, method = "random", seed = 1)
head(candidates)
#>       site_id        x        y      type       dem     slope      ndvi
#> 1 candidate_1 265.5087 814.2518 candidate 116.55597 6.7796984 0.4365338
#> 2 candidate_2 372.1239 928.7772 candidate 114.80089 1.0778848 0.6631923
#> 3 candidate_3 572.8534 147.4810 candidate  87.55092 3.9554009 0.1469143
#> 4 candidate_4 908.2078 749.8217 candidate  91.51938 8.1660975 0.6208423
#> 5 candidate_5 201.6819 975.6573 candidate 119.62574 0.7472459 0.5826633
#> 6 candidate_6 898.3897 974.7925 candidate  93.30237 0.4474288 0.3751245

method = "systematic" generates a regular grid instead of random points — useful when you want even coverage of the candidate pool itself.

Excluding Candidates Near the Inaccessible Sites

filtered <- ss_alt_filter_buffer(candidates, inaccessible, min_distance = 100)
cat("candidates before:", nrow(candidates), " after buffer:", nrow(filtered), "\n")
#> candidates before: 300  after buffer: 271

Computing Similarity for One Target Site

env_vars <- c("dem", "slope", "ndvi")
target_values <- unlist(terra::extract(r, inaccessible[1, c("x", "y")], ID = FALSE))

scores <- ss_alt_similarity(target_values, filtered[env_vars], method = "mahalanobis")
summary(scores)
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#>  0.3048  0.4426  0.5072  0.5346  0.6010  0.9345

Ranking Alternatives

top3 <- ss_alt_rank(scores, filtered, n_select = 3, target_site_id = "site_1")
top3[, c("site_id", "target_site_id", "similarity_score", "similarity_rank")]
#>           site_id target_site_id similarity_score similarity_rank
#> 90   candidate_90         site_1        0.9344683               1
#> 232 candidate_232         site_1        0.9344683               2
#> 193 candidate_193         site_1        0.9311112               3

Standardizing Raw Site Data

Field data often arrives as a plain CSV with arbitrary column names. ss_alt_standardize_sites() normalizes it to the site_id/x/y/type layout the other ss_alt_* functions expect:

raw_sites <- data.frame(lon = c(200, 500), lat = c(200, 500), Point_ID = c("A1", "A2"))
ss_alt_standardize_sites(raw_sites, x_col = "lon", y_col = "lat", id_col = "Point_ID")
#>   site_id   x   y         type
#> 1      A1 200 200 inaccessible
#> 2      A2 500 500 inaccessible

Choosing Parameters

method: use "mahalanobis" when covariates are correlated and the candidate pool is reasonably large; fall back to "euclidean" for speed on very large problems, or "gower" when covariates mix continuous and categorical variables.
n_candidates: larger pools give better matches but cost more to generate and compare. ss_alt_sites() defaults to max(n_alternatives * 50, 1000).
min_distance_buffer: set to a distance that meaningfully separates alternatives from the original inaccessible areas (e.g. the size of the inaccessible zone itself). Leave NULL if spatial proximity to the original site isn’t a concern.
Small candidate pools: with very few candidates relative to the number of covariates, the Mahalanobis covariance matrix can become poorly conditioned (all similarity scores end up nearly identical). Increase n_candidates if you see this.

Practical Workflow

library(soilsampling)
library(terra)

predictors <- rast("data/predictors.tif")
inaccessible <- read.csv("data/inaccessible_sites.csv")  # needs x, y columns

result <- ss_alt_sites(
  predictors,
  inaccessible,
  method = "mahalanobis",
  n_alternatives = 3,
  min_distance_buffer = 300,
  seed = 123,
  output_dir = "outputs"
)

result$alternatives

# outputs/ now contains:
#   similarity_analysis_sites.csv, similarity_analysis_inaccessible_sites.csv

When to Use Alternative Site Selection

✅ Use it when:

Some planned sites turn out to be inaccessible after the design was finalized
You want replacements that preserve the environmental coverage of the original design
You need documented, reproducible justification for the substitution (similarity scores and ranks)

❌ Don’t rely on it when:

The inaccessible sites are a small fraction of the design and simple nearest-neighbor substitution is sufficient
No environmental covariates are available to compare candidates against

References

Malone, B.P., Minasny, B., and Brungard, C. (2019). Some methods to improve the utility of conditioned Latin hypercube sampling. PeerJ 7, e6451. DOI: 10.7717/peerj.6451

soilsampling Package Authors

2026-08-02