Reproducing the Spatial Delta Model for Accessibility

1 Method Overview & Reproduction Scope

1.1 Core Idea

Two neighbourhoods can have the same amount of parkland nearby and be entirely different places to live. In one the park is across the road, uncrowded, easy to walk to. In the other the same hectares sit behind a motorway, shared with ten thousand people. Counting the greenspace cannot tell them apart.

The Spatial Delta Model takes that seriously by measuring both quantities and studying their difference. Access (α) is how much greenspace is reachable within a walking time — a count of supply. Accessibility (β) is how easy it is to actually reach and use, computed with a floating catchment model that discounts supply by travel distance and shares it among everyone competing for it. Both are standardised, and the model's subject is

The sign is a diagnosis. Where δ > 0 the greenspace that exists is easy to reach and lightly shared — the environment is doing more than its raw endowment suggests. Where δ < 0 the greenspace is there on paper but hard to use. Neither is visible from α or β alone.

• Research problem: access and accessibility are routinely treated as the same thing, so places where supply exists but cannot be used look identical to places where it can.
• One-line contribution: SDM standardises both measures onto a common scale, subtracts them, and turns the residual into a mapped, classifiable, explainable quantity.
• Why it matters: the gap is not noise. In the demo below it is spatially organised, it reverses with the walking distance considered, and roughly a quarter of it is explained by measurable features of the urban fabric.

Published reference · SDM

1.2 Method Logic

SDM runs in four movements: measure the two quantities, difference them, classify the result, and explain it.

Key concepts and equations

• Access (α) — the total greenspace area within the walking distance d₀:
  α_i = Σ_{j : dij ≤ d₀} G_j
  Supply, and nothing else. Two blocks with the same total score the same however that supply is arranged or however many people share it.
• Accessibility (β) — a Modified Gaussian Two-Step Floating Catchment Area:
  β_i = Σ_{j : dij ≤ d₀} G_j f(d_ij)² / Σ_{k : djk ≤ d₀} P_k f(d_kj)
  Each greenspace is discounted by how far away it is and divided among the population that can also reach it, so crowding and effort both enter. The decay f is Gaussian and truncated at d₀.
• The difference (δ) — both standardised, then subtracted. Standardising is what makes the subtraction meaningful: α is an area and β is a ratio, so they are only comparable once each is expressed in standard deviations of its own distribution.
• Spatial scale — everything is computed at six walking times (3–30 minutes, at 5 km/h), because "within reach" is a choice, not a fact, and the answer changes with it.
• Explaining δ — each candidate variable is cut into spatial strata by a regression tree, and the geographical detector scores how much of δ those strata account for (the Power of Determinant). Every variable appears twice: as its own value, and in a contextualised form carrying its local spatial pattern, so the analysis can say whether what matters is the quantity at a place or its arrangement around it. The detector is the same machinery as the OPGD tutorial.

Published reference · SDM

1.3 Reproduction Scope

Everything tagged Demo run regenerates from this folder; everything tagged Published reference is a static image from the article. It runs end to end with R 4.6.0.

How to use this page

• To reproduce the demo: Rscript run-all.R test, then Rscript run-all.R.
• To apply SDM to your own city: §2.3 gives the data contract, §2.4 the network-distance step, and §4.1 the porting checklist.
• To understand the detector: the driver analysis in Step 4 uses the same machinery as the OPGD tutorial, which covers it in depth.

2 Setup, Structure & Data Contract

2.1 Environment & Dependencies

Four packages carry the pipeline. A fifth is needed only for the Step 2 demonstration, and the step reports and continues without it.

# The pipeline
install.packages(c("sf", "rpart", "GD", "ggplot2"))

# Only for the accessibility demonstration in Step 2. This one lives on GitHub.
install.packages("remotes")
remotes::install_github("uva-bi-sdad/catchment")

2.2 Project Structure & Configuration

SDM/
├── sdm.html                        # this guide
├── run-all.R                       # entry point: full run, single steps, figures, test
├── assets/                         # figures shown in this guide (web copies)
├── config/project-config.R         # the ONLY file to edit for a new city
├── R/
│   ├── 00-config.R, 01-helpers.R   # paths, IO, column conventions
│   ├── 10-prepare-data.R           # read the table and the geometry, check the identity
│   ├── 20-access-accessibility.R   # how alpha and beta are produced (demonstration)
│   ├── 30-delta.R                  # delta, the green-city classification, by region
│   ├── 40-drivers-gozh.R           # Power of Determinant at every walking time
│   ├── 50-interactions.R           # pairwise interaction detector
│   ├── 90-tables.R                 # LaTeX tables + session info
│   └── p01 … p03                   # figure scripts
├── data/
│   ├── mbvars.csv                  # 8,259 blocks: alpha, beta, delta + 12 variables
│   ├── blocks.gpkg                 # 9,449 polygons: 8,276 residential, 1,173 parkland
│   ├── regions.gpkg                # the nine SA3 regions
│   └── provenance.md               # where every file came from
├── tests/                          # test-01 the published table; test-02 model properties
├── env/                            # requirements.md, session-info.txt
├── results/  tables/  figs/        # generated output, never hand-edited

Main configuration file

ANALYSIS_FILE <- "data/mbvars.csv"   # alpha, beta, delta + explanatory variables
BLOCKS_FILE   <- "data/blocks.gpkg"  # consumers and providers, as polygons
ID_COL        <- "MB_CODE21"         # joins the table to the geometry
REGION_COL    <- "SA3"

# "Within reach" is a choice, so the model is evaluated at six of them.
WALK_TIMES     <- c(3, 5, 10, 15, 20, 30)            # minutes
WALK_DISTANCES <- c(250, 417, 833, 1250, 1667, 2500) # metres, at 5 km/h

ACCESS_PREFIX        <- "u"    # alpha
ACCESSIBILITY_PREFIX <- "v"    # beta
DELTA_PREFIX         <- "df"   # delta = beta - alpha

# Six variables, each also in a contextualised form carrying its local pattern
VARS_RAW <- c("popdenskm2","dwedenskm2","shapefacto",
              "compactrat","neargsdist","neargsarea")
VARS_CTX <- c("lisav1","lisav2","lisav3","lisav4","lisav5","lisav6")

TREE_MINBUCKET <- 10           # smallest terminal node a stratifying tree may create

2.3 Data Contract

SDM needs two things: a table with the two measures at each spatial scale, and the geometry those measures were computed on.

2.4 Network Distances: the Manual Step

One part of the published workflow is not in the pipeline, and cannot be: the distances behind the published β are measured along the road network, not in straight lines, and the network was built by hand in QGIS. It matters — a park two hundred metres away across a river is not two hundred metres away — and it is why the demonstration in Step 2 does not reproduce the published columns.

The accompanying tutorial documents the procedure. In outline:

1. Get the road network. In QGIS, install the QuickOSM plugin (Plugins → Manage and Install Plugins), query OpenStreetMap for the highway layer over the study area, and export it as a shapefile.
2. Build the origins and destinations. Residential block centroids are the origins, greenspace centroids the destinations.
3. Run the network analysis. QGIS's routing tool takes the road layer, the origins and the destinations, and returns a matrix of shortest-path distances along the network.
4. Bring it back into R. Export the result and pass it to catchment_ratio() as the cost argument, in place of the straight-line distances the function computes by default.

Published reference · tutorial

Published reference · tutorial

3 Reproduction Pipeline, Results & Validation

3.1 Pipeline Execution

cd SDM                     # this folder
Rscript run-all.R test     # 1. check the published table is intact
Rscript run-all.R          # 2. full pipeline

Rscript run-all.R 30 40    # re-run single steps (10 20 30 40 50 90)
Rscript run-all.R figures  # only the figure scripts

3.2 Reproduce the Published Values

Three separate claims are checked, and they are not equally strong. Being clear about which is which is the point of this section.

The published table is intact — checked

The α, β and δ values are the authors' own. What has to hold is that they survived the trip into this folder and that the model's defining identity holds on them: δ = β − α at all six scales, to 10⁻⁶ (the rounding of the shipped CSV), on all 8,259 blocks, every one of which still matches its polygon.

Demo run

Published reference · SDM

The published findings are recovered — checked

The regional pattern falls out of the analysis rather than being asserted. At a 30-minute walk the demo puts Perth City lowest at δ = −0.765 and Fremantle highest at +0.412 — the two extremes of the paper's Fig. 5, at the values its axis shows (reproduced in §1.1).

Demo run

The measurement of α and β is demonstrated, not reproduced — stated plainly

Step 2 recomputes both measures from the block geometry. Against the published columns, access correlates at r = 0.61 and accessibility at r = −0.08. The reason is §2.4: the published β is built on road-network distances, and the demonstration uses straight lines. The demonstration is faithful to the practical tutorial — it reproduces that document's printed median of 0.0105 — but it is not the paper's β, and nothing downstream uses it.

3.3 Core Analytical Steps

Five stages: purpose → code → output → result → how to read it.

d <- read_analysis()

# The identity the whole model rests on: delta must be beta - alpha
# at every walking time, or nothing downstream means what it says.
worst <- 0
for (t in WALK_TIMES) {
  dev   <- max(abs(d[[col_delta(t)]] -
                   (d[[col_accessibility(t)]] - d[[col_access(t)]])))
  worst <- max(worst, dev)
}
if (worst > 1e-4) log_warn("delta does not match beta - alpha; check the input table")

# access: the greenspace area within the walking distance
D <- sf::st_distance(consumer_centroids, provider_centroids)
alpha <- as.numeric((D <= d0) %*% providers$AREASQKM21)

# accessibility: MG2SFCA — supply discounted by distance, shared among demand
beta <- catchment::catchment_ratio(
  consumers = consumer_centroids, providers = provider_centroids,
  weight = "gaussian", consumers_value = "Person",
  providers_value = "AREASQKM21", max_cost = d0, return_type = "region")

Step 3 · the model

The delta, and the green-city classification

What this step does

Differences the two standardised measures at each walking time, classifies every block by the sign of the result, and summarises by region. The near-zero class needs a band rather than an exact value, since δ is continuous: one tenth of a standard deviation, which keeps the classification comparable across scales.

Code

R — from R/30-delta.R Copy

a <- d[[col_access(t)]]; b <- d[[col_accessibility(t)]]; dl <- d[[col_delta(t)]]

# Summarise the difference at each walking time
data.frame(walk_minutes = t,
           mean_delta     = mean(dl),
           share_positive = mean(dl > 0),
           share_negative = mean(dl < 0),
           cor_access_accessibility = stats::cor(a, b))

# The green-city classification: a band of width sd/10 around zero is "zero"
eps <- stats::sd(dl) / 10
k   <- ifelse(dl > eps, "positive", ifelse(dl < -eps, "negative", "zero"))

Example output

Walk	Distance	δ > 0	δ ≈ 0	δ < 0	cor(α, β)
3 min	250 m	55.7%	8.3%	36.0%	+0.749
5 min	417 m	52.1%	10.9%	37.0%	+0.716
10 min	833 m	51.1%	10.6%	38.3%	+0.745
15 min	1250 m	49.1%	11.9%	39.0%	+0.709
20 min	1667 m	47.2%	10.4%	42.4%	+0.711
30 min	2500 m	43.2%	9.6%	47.2%	+0.680

Demo run

Published reference · SDM

Demo run

◆How to read this result

• The crossover is the headline. Positive blocks fall from 55.7% to 43.2% as the walk lengthens, and negative ones rise past them. Close to home, Perth's greenspace is easier to use than its bare quantity suggests; over a half-hour walk the opposite holds. A single-distance study would have found one of these and missed that it reverses.
• α and β correlate at about +0.7, not +1. They are measuring related things — which is why differencing them is meaningful rather than trivial — but roughly half the variance is not shared, and that half is what δ isolates.
• Perth City is the clearest case. Lowest δ at every scale, reaching −0.765 at thirty minutes: the city centre has greenspace, but it is distant, contested and hard to reach relative to how much is nominally there.
• The near-zero band is a choice, so it is stated. One tenth of a standard deviation; about a tenth of blocks land in it. Widen it and the story becomes "mostly neutral", narrow it and the neutral class vanishes — report whichever you use.

Step 4

What explains the difference

What this step does

Each variable is cut into spatial strata by a regression tree, and the geographical detector scores how much of δ those strata explain. Every variable is tested twice: as its own value, and in a contextualised form carrying its local spatial pattern.

Code

R — from R/40-drivers-gozh.R Copy

# the variable is continuous, so a tree turns it into spatial strata
tr <- rpart::rpart(df ~ v, data = dd, minbucket = TREE_MINBUCKET)
dd$strata <- as.character(as.numeric(tr$where))

# then the detector scores how much of delta those strata account for
g <- GD::gd(df ~ strata, dd)      # g$Factor$qv is the Power of Determinant

Example output

Demo run

Published reference · SDM

Walk	Strongest single variable	PD	All variables jointly
3 min	Distance to near greenspace (raw)	0.116	0.262
5 min	Distance to near greenspace (contextualised)	0.116	0.235
10 min	Area of near greenspace (raw)	0.055	0.197
15 min	Area of near greenspace (contextualised)	0.074	0.272
20 min	Area of near greenspace (raw)	0.137	0.250
30 min	Region	0.216	0.426

◆How to read this result

• What drives the gap changes with the distance considered. At a three-minute walk it is how far the nearest park is. By thirty minutes the strongest single predictor is simply which region you are in (PD 0.216) — at that range the difference is a property of urban structure at the district scale, not of the individual block.
• Arrangement often beats quantity. The contextualised forms outscore their raw counterparts at four of the six scales. How greenspace and population are patterned around a block explains the access–accessibility gap better than how much of each there is — which is what the difference was built to capture.
• Together they explain a fifth to two fifths. Jointly, PD runs 0.197 to 0.426, highest at thirty minutes. Most of δ remains unexplained by these twelve variables, and reporting that honestly is better than implying the gap is solved.
• Region is weak at short range and dominant at long. PD 0.013 at five minutes against 0.216 at thirty: administrative geography says almost nothing about your walk to the nearest park, and a great deal about your access to the city's greenspace as a whole.

Step 5

Interactions — pairs that explain more together

What this step does

Overlays each pair of stratifications and scores the combined zones, classifying the pair into the five standard interaction types.

Code

R — from R/50-interactions.R Copy

# Each variable becomes spatial strata through a tree fitted against delta
strata_of <- function(x) {
  if (is.character(x) || is.factor(x)) return(as.character(x))
  tr <- rpart::rpart(df ~ v, data = data.frame(df = y, v = x),
                     minbucket = TREE_MINBUCKET)
  as.character(as.numeric(tr$where))
}

# Power of Determinant of one stratification, and of a pair combined
pd_strata <- function(s)
  as.numeric(GD::gd(df ~ strata, data.frame(df = y, strata = s))$Factor$qv)

single <- vapply(lapply(vars, strata_of), pd_strata, numeric(1))

Example output

Variable 1	Variable 2	PD₁	PD₂	PID	Type
Area of near greenspace	Contextualised population density	0.039	0.072	0.282	Enhance, nonlinear
Region	Contextualised distance to near greenspace	0.013	0.116	0.248	Enhance, nonlinear
Contextualised dwelling density	Contextualised distance to near greenspace	0.083	0.116	0.208	Enhance, nonlinear

Demo run

◆How to read this result

• The best pair more than doubles the best single. PID 0.282 against a strongest single PD of 0.116 — a gain of 144%. Greenspace area and population pattern say far more about the gap jointly than either says alone, which is exactly what you would expect of a measure built from supply and competition.
• Nonlinear-enhance dominates: 28 of 45 pairs. The joint effect exceeds the sum of the parts. These variables are not independent contributions to be added up; they condition each other.
• Region is weak alone but strong in company. PD 0.013 by itself, PID 0.248 paired with contextualised distance — the district matters, but only once you know the local greenspace geography within it.

3.4 Reading the Results Together

Published reference · SDM

4 Adaptation, Writing & Reproducibility

4.1 Port to Your Domain

Nothing in the model is specific to greenspace. It needs a service that people travel to, a population that travels to it, and a way of measuring the journey. Swap those three and the method is unchanged:

What to edit

1. config/project-config.R — the three data paths, ID_COL and REGION_COL, and the consumer and provider categories with their demand and supply fields.
2. WALK_TIMES and WALK_DISTANCES — the scales that matter for your service. A clinic is not a park: the relevant range may be a drive rather than a walk, in which case set the distances from the appropriate speed and say so.
3. VARS_RAW and VARS_CTX — the candidate explanations, in both plain and contextualised form if you have them.
4. Compute α and β for your own case (Step 2 shows the shape of it; §2.4 covers the network distances), write them into the analysis table under the configured prefixes, and the rest of the pipeline runs unchanged.

4.2 Write the Paper

Manuscript skeleton

1  Introduction        — access is not accessibility; why the gap matters;
                          the contribution (3 aims)
2  Spatial Delta Model — 2.1 Access (alpha): supply within reach
                          2.2 Accessibility (beta): MG2SFCA, all formulas here
                          2.3 The difference (delta) and its interpretation
                          2.4 Spatial scales, and the distance basis
3  Case & data         — study area, consumers, providers, explanatory variables
4  Results             — 4.1 access and accessibility mapped
                          4.2 the difference, and how it changes with distance
                          4.3 the green-city classification by region
                          4.4 what explains the difference (PD by scale)
                          4.5 interactions between explanations
5  Discussion          — high-alpha vs low-alpha negatives; the scale reversal;
                          what remains unexplained; distance-basis limitations
6  Conclusions         — 6-8 sentences answering the aims

Results-to-manuscript map

4.3 Final Reproducibility Package

References & credits

Demo outputs were generated by the scripts in this folder from the authors' own analysis table. Reference figures are reproduced from the article, which is open access under CC BY, and from the accompanying practical tutorial.

1. Lin G, Song Y, Xu D, Swapan MSH, Wu P, Hou W, Xiao Z (2024). Interpreting differences in access and accessibility to urban greenspace through geospatial analysis. International Journal of Applied Earth Observation and Geoinformation 129:103823. doi:10.1016/j.jag.2024.103823 · PDF (source of the method, the case data, and the Fig. 1, 3, 4, 5, 6 and 7 reference images)
2. Delamater PL (2013). Spatial accessibility in suboptimally configured healthcare systems: a modified two-step floating catchment area (M2SFCA) metric. Health & Place 24:30–43. doi:10.1016/j.healthplace.2013.07.012 (the accessibility model behind β)
3. Song Y, Wang J, Ge Y, Xu C (2020). An optimal parameters-based geographical detector model. GIScience & Remote Sensing 57(5):593–610. doi:10.1080/15481603.2020.1760434 · OPGD tutorial (the detector used in Steps 4 and 5)
4. The catchment R package (Social and Decision Analytics Division, University of Virginia) — used for the MG2SFCA demonstration in Step 2.