Reproducing the Degree of Spatial Interpretability (DSI)

1 Method Overview & Reproduction Scope

1.1 Core Idea

Conventional model validation asks how accurate a model is — R², RMSE, MAE. The degree of spatial interpretability (DSI) asks a different question that accuracy metrics cannot answer: how much of the spatial structure present in the data has the model actually explained, rather than passed through into its residuals? A model can lower its errors by smoothing toward the local mean while leaving the spatial arrangement of those errors intact — high accuracy, low spatial interpretability.

The idea rests on a principle from spatial statistics: any spatial structure in the response that a model fails to capture is transferred to its residuals. Measuring a spatial characteristic on the response field and again on the residual field, the relative decrease tells you how much the model absorbed. DSI turns that into a number; its local extension, the degree of geocomplexity (DG), turns it into a map.

• Research problem: model selection for spatial prediction is dominated by accuracy metrics that treat model error as spatially uniform, so a model that reproduces the spatial structure of the response and a model that merely interpolates it are scored alike.
• One-line contribution: DSI measures the residual-based attenuation of spatial autocorrelation (Moran's I) and spatial heterogeneity (geographical-detector Q) from the response to each model's residuals, combines them into a θ range per model, and — via DG — localises the same attenuation logic point by point.
• Why it matters: applied to a set of competing models, accuracy and DSI often select different winners; in the demo below the best-R² model (random forest) ranks only fourth of five on θ_max, while xgbTree leads interpretability.

1.2 Method Logic

DSI is built from three measurements, each applied twice — once to the response field, once to a model's residual field. Two are global (autocorrelation, heterogeneity); one is local (geocomplexity, used by DG). The method composes four earlier building blocks:

Published reference · DSI

Key concepts and equations

Three equations drive the whole pipeline:

• Individual characteristic η (Eq. 1):
  η = ( δ₀ − δ_r ) / δ₀
  δ₀ is the characteristic measured on the dependent variable; δ_r the same characteristic on the residuals. η → 1 means the residual field keeps almost none of the structure (the model absorbed it); η = 0 means it absorbed none; η < 0 means the model amplified structure that was not in the data. Autocorrelation is measured by Moran's I (KNN weights, projected CRS in metres); heterogeneity by the geographical-detector Q statistic, with strata cut from the predictors — Wang et al. (2010) doi.
• Multiple characteristics θ (Eq. 2):
  θ_min = min(η₁,…,η_n)  θ_probable = max(η₁,…,η_n)  θ_max = 1 − ∏(1 − η_i)
  Across several characteristics the result is a range, not a single number: θ_min reads the components as shared, θ_max as independent contributions. The width between them is itself diagnostic — narrow means the model treats the characteristics alike; wide means it is strong on one and weak on another — Liu, Song & Yi (2026) PDF.
• Local extension DG (Eq. 3):
  DG_i = ( G_0,i − G_r,i ) / G_0,i
  G is the geocomplexity of Zhang et al. (2023) (PDF), computed at every location on the response and on each residual field. DG answers where a model explains local structure — the basis for a map of which model is locally best. The ratio is unbounded and heavy-tailed (G in the denominator), so summaries report the median alongside the mean — Liu et al. (2026) PDF.

1.3 Reproduction Scope

Everything tagged Demo run on this page regenerates from the scripts and data in this folder; everything tagged Published reference is shown as a static comparison image or table from the two source articles. The demo runs end to end with R 4.6.0, and the built-in test reproduces every published value of the DSI paper's Table 3 (row lm) within 0.005.

Published applications of this exact method

How to use this page

• To reproduce the demo: this folder already contains everything — the R scripts (R/), the configuration (config/project-config.R), the paper's demo dataset (data/reference/), and the verification tests (tests/). Run Rscript run-all.R test, then Rscript run-all.R; every expected output is shown on this page so you can verify each step.
• To apply DSI to your own problem: read §2.3 (data contract), point INPUT_FILE in the §2.2 config at your CSV, then run the same scripts unchanged (§4.1).
• To write the paper: §4.2 gives the section skeleton pre-structured for a DSI application and two worked writing examples with the demo numbers filled in.

2 Setup, Structure & Data Contract

2.1 Environment & Dependencies

The pipeline is pure R and runs from one entry point. Required packages stop the run if missing; optional packages only disable the model or module that needs them, and the pipeline reports what it skipped — so a run always produces output. Install everything from the R console:

# Required — the pipeline stops without these
install.packages(c("sf", "spdep", "rpart"))

# Optional — the local DG module and the figures
install.packages(c("geocomplexity", "ggplot2", "scales", "patchwork"))

# Optional — the full model set of the demo configuration
install.packages(c("ranger", "xgboost", "Cubist", "FNN",
                   "pls", "earth", "kernlab", "gbm"))

2.2 Project Structure & Configuration

One entry script drives everything: run-all.R runs the six pipeline steps, the six figure scripts, or the verification tests, and every step writes to data/derived/, results/, tables/ and figs/. Each step reads what the previous one wrote, so a single step can be re-run after a config change without repeating the whole pipeline. Deleting the four output folders and re-running regenerates everything shown on this page.

DSI/
├── dsi.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 domain
├── R/
│   ├── 00-config.R … 05-models.R  # library layer: config, helpers, metrics,
│   │                              #   DSI, DG, model registry
│   ├── 09-make-synthetic-data.R   # built-in demo generator (INPUT_FILE = NULL)
│   ├── 10 … 60                    # pipeline: data, models, DSI, DG,
│   │                              #   sensitivity, tables
│   └── p01 … p06                  # figure scripts
├── data/
│   ├── raw/                       # your input CSV goes here
│   ├── reference/                 # the DSI paper's demo dataset + original code
│   └── derived/                   # generated: analysis data, split, residuals
├── tests/                         # test-01 reproduces the published values;
│                                  #   test-02 checks 17 metric properties
├── env/                           # requirements.md, session-info.txt
├── results/  tables/  figs/       # generated output, never hand-edited

Main configuration file

INPUT_FILE    <- NULL          # NULL = synthetic demo; else "data/raw/your.csv"
RESPONSE      <- "y"           # name of the dependent-variable column
PREDICTORS    <- NULL          # NULL = every column except lon/lat/response
DATA_CRS      <- 4326          # CRS of the lon/lat columns
PROJECTED_CRS <- 3577          # a projected CRS in METRES (not degrees!)
K_MORAN       <- 8             # KNN neighbours for Moran's I (6-10 dense, 10-20 sparse)
K_GEOC        <- 18            # neighbourhood for geocomplexity (DG)
K_SENSITIVITY <- c(6, 8, 10, 14, 18, 22)   # k values re-run in Step 5
STRATIFY_METHOD <- "tree"      # "tree" (paper) or "kmeans" — never the response
MODELS <- c("rpart","rf","xgbTree","cubist","knn",
            "pls","earth","svmRadial","gbm")
TRAIN_FRACTION <- 0.7          # 70/30 split, as in both papers
SEED           <- 42

Key parameters and where they come from

2.3 Data Contract

The pipeline consumes one table, one row per sampling location. Everything except the coordinates and the response is treated as a predictor — used both to fit models and to define the Q-value strata. Units can be anything point-referenced: grid cells (DSI application), sample plots (DG application), stations, polygon centroids.

Example analysis-ready table

The generator behind the demo (or: how to build a controlled test)

When INPUT_FILE is NULL, the pipeline builds a synthetic field with known structure — a smooth autocorrelated component, a three-region regime term, and a spatially correlated error. Because the truth is known, a model that recovers it should score high DSI and a model that ignores space should not. The same generator is the starting point for the simulation section of a methods paper.

# Smooth spatial field from an exponential covariance (autocorrelation)
smooth_field <- function(range_km, sd = 1) {
  covm <- sd^2 * exp(-d_km / range_km)
  L <- chol(covm + diag(1e-6, n))
  as.numeric(t(L) %*% rnorm(n))
}
# Regime term: three east-west bands (heterogeneity -> picked up by Q)
regime_effect <- c(-4, 0, 5)[cut(grid$lon, breaks = 3, labels = FALSE)]
# Spatially correlated error, so a good aspatial fit still leaves structure
spatial_error <- 2.2 * smooth_field(range_km * 0.5, 1)

y <- 12 + 0.9*(temperature - mean(temperature)) + 1.5*(soil_ph - mean(soil_ph)) +
     regime_effect + spatial_error + rnorm(n, sd = 1)

3 Reproduction Pipeline, Results & Validation

3.1 Pipeline Execution

Six pipeline steps, six figure scripts and two verification tests run from one entry point. Each step writes to data/derived/, results/, tables/ or figs/ and reads what the previous step wrote, so any step can be re-run in isolation after a config change.

cd DSI                    # this folder
Rscript run-all.R test    # 1. verify the implementation reproduces the paper
Rscript run-all.R         # 2. full pipeline on the built-in synthetic data

Rscript run-all.R 30 40   # re-run single steps (10 20 30 40 50 60)
Rscript run-all.R figures # only the figure scripts

3.2 Reproduce the Published Values

Before trusting any number on your own data, confirm the code reproduces the published method. run-all.R test runs the template's metric functions on the demo dataset released with the DSI paper (data/reference/dsi-paper-demo-dataset.csv, 586 test points) and compares nine values against Table 3 (row lm) and Figure 6 of Liu, Song & Yi (2026).

q_statistic <- function(x, strata) {
  strata <- as.factor(strata)
  keep <- is.finite(x) & !is.na(strata)
  x <- x[keep]; strata <- droplevels(strata[keep])
  sst <- sum((x - mean(x))^2)               # total sum of squares
  if (sst <= 0) return(NA_real_)
  ssw <- sum(vapply(split(x, strata),       # within-stratum sum of squares
                    function(v) sum((v - mean(v))^2), numeric(1)))
  1 - ssw / sst
}

3.3 Core Analytical Steps

Five stages, each with the same anatomy: purpose → code → output → result → how to read it. Steps are numbered 1–5 on this page while the script files keep their own numbers (Step 1 runs R/10-prepare-data.R, Step 3 runs R/30-compute-dsi.R, and so on) — R/00–05 is the shared library layer.

Step 1

Prepare data — raw input → analysis-ready table

What this step does

Reads the input (or generates the synthetic demo when INPUT_FILE is NULL), drops rows with missing coordinates/response and duplicate coordinates, optionally log-transforms the response, reports predictor pairs with |r| > 0.8 without acting on them, and writes data/derived/analysis-data.csv. Everything downstream reads only the derived file, so the raw data is never modified.

Code

terminal Copy

Rscript run-all.R 10

Example output

== Step 1/6 Prepare data ====================================== INPUT_FILE is NULL; generating the synthetic demonstration dataset 1024 rows, 9 columns response: y 6 predictor(s): temperature, precipitation, elevation, soil_ph, distance_km, noise_var no predictor pair exceeds |r| = 0.8 wrote results/variable-summary.csv (7 rows) wrote data/derived/analysis-data.csv (1024 rows) analysis set: 1024 points, 6 predictors

Demo run

Paper reference

In Liu, Song & Yi (2026) this step aggregates 115,083 Australian plant-diversity plots to 1,958 0.5°×0.5° grid cells (richness as the response, 19 abiotic predictors — PDF). The synthetic demo mirrors that structure at smaller scale so the pipeline runs without downloading the case data.

◆How to read this result

• Confirm there is structure to explain. DSI is an attenuation ratio; if the response carries no spatial pattern, every η is noise over noise. A visibly patchy response map (like this one) is the precondition for the whole analysis.
• Two kinds of structure, deliberately. The gradient (heterogeneity) is what the Q statistic detects; the local clustering (autocorrelation) is what Moran's I detects. A case that shows only one makes only one η interpretable.
• Collinearity is reported, not fixed. The step flags |r| > 0.8 pairs so you decide; because dropping a predictor reshapes the Q-value strata, that decision belongs in the manuscript, not in a silent default.

✓Checkpoint

data/derived/analysis-data.csv exists with 1,024 rows and 6 predictors, and the console reports no predictor pair above |r| = 0.8, before you move on.

Step 2

Fit models & accuracy — predictions, residuals, the conventional ranking

What this step does

Splits the data 70/30 by quantiles of the response, fits every available model from the configured set, and computes residuals on the held-out test set. Accuracy (R², RMSE, MAE) is the conventional ranking that DSI will be contrasted against.

!Residuals must come from held-out data

Training residuals of an overfitted model approach zero, which drives η toward a spurious 1. Both the paper and this template evaluate on the test set; the split is stratified so the test set keeps the response distribution and therefore a comparable δ₀.

Code

terminal Copy

Rscript run-all.R 20

Example output

== Step 2/6 Fit models and extract residuals ================== stratified split: 714 training / 310 test (70% / 30%) ! skipped (package not installed): pls [pls], earth [earth], svmRadial [kernlab], gbm [gbm] fitting 5 model(s): rpart, rf, xgbTree, cubist, knn wrote data/derived/test-residuals.csv (310 rows) wrote results/accuracy-metrics.csv (5 rows) best R2: rf (0.366) | lowest RMSE: rf (4.122)

Model	R²	RMSE	MAE
rf	0.366	4.122	3.383
xgbTree	0.318	4.407	3.530
knn	0.309	4.320	3.557
cubist	0.309	4.355	3.479
rpart	0.217	4.779	3.879

Demo run

Paper reference

In the DSI paper's Australian case, rf and cubist achieved the best accuracy (R² ≈ 0.48–0.49), with svmRadial, gbm, gam, xgbTree close behind and earth, lm, knn, pls weakest — a ranking the DSI analysis then partly overturns (PDF).

◆How to read this result

• This is the baseline you will challenge, not the answer. The accuracy table establishes the conventional hierarchy; the entire point of DSI is to show that hierarchy is incomplete.
• Small R² gaps hide large behavioural differences. Four models sit within 0.31–0.37 R² — a reviewer would call them "comparable." Step 3 shows their θ_max ranges from 0.47 to 0.74, a difference accuracy cannot see.
• Deeper thinking: the modest R² (≈ 0.37) is expected when predictors are purely abiotic and a spatially correlated error is present by construction — exactly the regime where spatially structured residuals, and therefore DSI, matter most.

✓Checkpoint

results/accuracy-metrics.csv holds 5 rows with rf best at R² = 0.366; models with missing packages are listed as skipped, not failed.

Step 3 · the core result

Compute DSI — degree of spatial interpretability per model

What this step does

Under one fixed set-up — the same neighbours, strata, and evaluation points for every model — it measures Moran's I and the Q statistic on the response and on each residual field, forms η for each characteristic (Eq. 1), and combines them into the θ range (Eq. 2).

Code

R/03-dsi.R — the indicator (called by step 30) Copy

dsi_eta <- function(delta_0, delta_r) {
  if (abs(delta_0) < 1e-8) return(NA_real_)     # nothing to explain
  (delta_0 - delta_r) / delta_0
}
dsi_theta <- function(etas) {
  etas <- etas[!is.na(etas)]
  list(theta_min      = min(etas),
       theta_probable = max(etas),
       theta_max      = 1 - prod(1 - etas))
}

terminal Copy

Rscript run-all.R 30

Example output

== Step 3/6 Compute DSI ======================================= Moran's I weights: 8-nearest neighbours, row-standardised, EPSG:3577 Q-value strata: tree method, 20 strata wrote results/dsi-metrics.csv (5 rows) No diagnostic flags raised. wrote results/accuracy-vs-dsi.csv (5 rows) Spearman rho between R2 and theta_probable: 0.800 accuracy and interpretability disagree: rf leads on R2, xgbTree on theta_probable

Model	δar	ηa	δhr	ηh	θmin	θprob	θmax
xgbTree	0.426	0.337	0.208	0.601	0.337	0.601	0.735
rpart	0.516	0.197	0.301	0.423	0.197	0.423	0.537
knn	0.538	0.163	0.296	0.433	0.163	0.433	0.526
rf	0.561	0.126	0.290	0.445	0.126	0.445	0.515
cubist	0.580	0.097	0.305	0.416	0.097	0.416	0.473

Demo run

Published reference · DSI

Paper reference — Table 3 (all models)

Model	δar	ηa	δhr	ηh
cubist	0.027	0.934	0.103	0.799
rf	0.039	0.905	0.127	0.752
xgbTree	0.048	0.883	0.095	0.815
svmRadial	0.059	0.856	0.157	0.694
gbm	0.078	0.809	0.122	0.762
gam	0.095	0.768	0.133	0.741
earth	0.142	0.653	0.164	0.680
knn	0.143	0.650	0.186	0.637
lm	0.189	0.538	0.203	0.604
pls	0.272	0.335	0.274	0.466

◆How to read this result

• η is an attenuation, not an error. η^a = 0.337 for xgbTree means it removed 34% of the response's autocorrelation from its residuals; η^h = 0.601 means it removed 60% of the heterogeneity. Higher is better.
• Negative η is a finding, not a bug — watch for it. η < 0 means residuals more spatially clustered than the response: the model manufactures pattern, the typical signature of a global linear or additive fit that cannot bend across a regime boundary. The pipeline flags any negative η in results/dsi-diagnostics.txt (none occurs in this five-model demo).
• Read the two components, not just θ. A model can reach a respectable θ_max from one strong component and one weak one. xgbTree here is far better on heterogeneity (0.60) than autocorrelation (0.34) — the wide θ span (0.34→0.74) says so at a glance.
• Demo vs paper — same logic, different winner. In the paper's real data every tree/ensemble model reaches η > 0.75 and cubist leads; on the harder synthetic case the absolute η are lower and xgbTree leads. The ranking mechanism reproduces even when the numbers differ, which is what a method port must demonstrate.
• Deeper thinking: that rf (best accuracy) sits mid-table on θ is the payload — it interpolates well within dense clusters yet leaves neighbourhood-scale residual structure. Accuracy rewarded the interpolation; DSI penalised the leftover structure.

✓Checkpoint

results/dsi-metrics.csv holds 5 rows; the console reports "No diagnostic flags raised" and ρ(R², θ_probable) = 0.800 with rf and xgbTree as the divergent pair.

Step 4 · optional module

Compute DG — point-wise geocomplexity & the locally optimal model

What this step does

Computes local geocomplexity of the response and of each residual field at every test location, forms DG point-by-point (Eq. 3), summarises the distribution, and assigns each location the model with the highest local DG — producing a map that argues no single model wins everywhere. Needs the geocomplexity package; set RUN_DG <- FALSE to skip.

Code

terminal Copy

Rscript run-all.R 40

Example output

== Step 4/6 Compute DG ======================================== geocomplexity neighbourhood: 18 nearest neighbours wrote results/dg-pointwise.csv (1550 rows) wrote results/dg-summary.csv (5 rows) wrote results/dg-optimal-model.csv (310 rows) locally best model: xgbTree at 32.6% of points (101 of 310) 5 model(s) are locally optimal somewhere

Model	Mean DG	Median DG	Share DG > 0
rpart	31.55	0.512	63.2%
xgbTree	22.36	0.504	64.8%
cubist	29.80	0.407	62.6%
knn	16.88	0.402	61.9%
rf	28.45	0.349	59.0%

Demo run

Published reference · DG

Paper reference

The DG paper reports mean DG values from −0.57 (lm) to 1.61 (rf) with an unbounded, heavy-tailed distribution — the reason this template reports the median beside the mean (PDF).

!Quote the median, not the mean

DG has local complexity in its denominator, so points where complexity is near zero produce extreme ratios. The demo means (17–32) are dominated by a handful of such points; the medians (0.35–0.51) are the honest summary. The pipeline warns when the mean and median disagree in sign.

◆How to read this result

• DG localises what DSI totals. DSI said xgbTree explains the most structure overall; DG shows it does so at most — but not all — locations, and names the second-best model elsewhere.
• "No single best model" is the practical claim. The map is the argument for local model selection: in a real study you might deploy different models in different regions, an option accuracy metrics never surface.
• Rank ≠ magnitude. A model can be locally best at many points while explaining little in absolute terms — always read the map together with the summary table before calling a model good.
• Deeper thinking: with a dense, near-regular synthetic grid the niches are diffuse; in the paper's clustered biomass plots they are sharp and geographic (coastal vs inland). Sparse or clustered sampling makes the map more decisive — and, below a few multiples of K_GEOC, noisier.

✓Checkpoint

results/dg-summary.csv holds 5 rows with medians between 0.349 and 0.512, and results/dg-optimal-model.csv assigns all 310 test points; the console reports xgbTree locally best at 32.6%.

Step 5

Sensitivity — is the result an artefact of the settings?

What this step does

Two settings are analyst choices, not properties of the data: the neighbourhood size k behind Moran's I, and the rule that defines the Q-value strata. This step recomputes DSI across k ∈ {6, 8, 10, 14, 18, 22} and under an alternative partition (k-means instead of tree), so the manuscript can state how much the ranking moves.

Code

terminal Copy

Rscript run-all.R 50

Example output

== Step 5/6 Sensitivity of DSI to measurement settings ======== wrote results/sensitivity-k.csv (30 rows) k varied over {6, 8, 10, 14, 18, 22}; largest change in theta_probable is 0.00% (cubist at k = 6) DSI is stable across neighbourhood sizes (all changes below 5%). wrote results/sensitivity-strata.csv (5 rows) model ranking agreement across stratification methods: rho = -0.900 wrote results/sensitivity-summary.txt

Demo run

Demo run

Paper reference

Section 5.2 of Liu, Song & Yi (2026) compares standard vs. spatial block cross-validation for rf and finds all DSI components move by less than 5% (Δθ_probable = 3.4%, Δθ_min = 4.8%, Δθ_max = 1.3%) — the basis for its robustness claim (PDF).

◆How to read this result

• Neighbourhood size is safe here; report it anyway. θ_probable is invariant to k for models where heterogeneity dominates (η^h does not use k), which is why the lines are flat. State the chosen k and its stability so a reviewer cannot suspect cherry-picking.
• Stratification is not neutral — this is the honest caveat. Tree and k-means strata rank the heterogeneity component differently (ρ = −0.90). Name the stratification rule in the Methods section and, when rankings differ, report both. Hiding this is the kind of fragility reviewers probe.
• Deeper thinking: the disagreement is larger on the demo than it would be in a strongly heterogeneous real case, because the synthetic regime term is only one of several signals. The more the response is genuinely regime-driven, the more any sensible partition agrees.

✓Checkpoint

results/sensitivity-summary.txt states both findings: stable across k (all changes below 5%) and partition-dependent across stratification rules (ρ = −0.900, report both rankings).

3.4 Validation & Interpretation

The whole project converges on one figure: interpretability against accuracy. Everything else is scaffolding for the moment where the two criteria are placed on the same axes and disagree.

Demo run

Published reference · DSI

Published reference · DG

Validation summary

4 Adaptation, Writing & Reproducibility

4.1 Port to Your Domain

The pipeline was designed to move between domains without touching the analytical scripts: the config declares the data, the preparation step builds the analysis table, and everything downstream is generic. Two published applications illustrate the span:

What to edit

1. config/project-config.R — point INPUT_FILE at your CSV, name the RESPONSE column, set PROJECTED_CRS to a metric CRS for your region, and revisit K_MORAN (6–10 dense samples, 10–20 sparse). The single adaptation point of §2.2.
2. Optionally list DROP_COLS (IDs, dates) and set LOG_TRANSFORM_RESPONSE <- TRUE for right-skewed responses (biomass, income).
3. Nothing else: run Rscript run-all.R test once (the published-value check does not depend on your data), then Rscript run-all.R, and read results/dsi-diagnostics.txt before drafting anything.

4.2 Write the Paper

The manuscript mirrors the pipeline: Methods sections describe the indicator and the experiment design 1:1 with the scripts, every Results sentence carries a number pulled from tables/*.tex or results/ so nothing is retyped, and the Discussion is built on the accuracy–interpretability divergence and the sensitivity sweeps. Section 2 (the indicator) is largely written already, because its definition does not change between domains — the writing work concentrates in the case study, the Results, and the mechanism of the divergence.

Manuscript skeleton

1  Introduction            — 6 paragraphs, narrowing: background → gap → aim   (~1,200 words)
2  The indicator           — 2.1 Concept (no formulas)                          (~1,400 words;
                             2.2 Calculation process (ALL formulas live here):   largely
                                 2.2.1 Spatial characteristics of the response   pre-written)
                                       (autocorrelation, heterogeneity,
                                       geocomplexity; significance δ0)
                                 2.2.2 Residual attenuation of each
                                       characteristic (η)
                                 2.2.3 θ aggregation (θ_min, θ_probable, θ_max)
                                 2.2.4 Local degree of interpretability (DG)
3  Case study              — study area, data, experiment design — no results  (~1,300 words)
4  Results                 — accuracy → DSI → divergence → local (DG) →        (~1,600 words)
                             sensitivity, mirroring the pipeline steps 1:1;
                             every finding sentence carries a number;
                             one interpretive takeaway per subsection
5  Discussion              — why the divergence happens, contribution,         (~1,200 words)
                             limitations; no subsection headings
6  Conclusions             — 6-7 sentences, each mapping to a contribution     (~300 words)

Results-to-manuscript map

4.3 Final Reproducibility Package

Everything an editor, reviewer, or reader needs sits in this folder; the three blocks below are what a submission's data-and-code availability statement should point to.

References & credits

Demo outputs were generated by the scripts in this folder. Reference figures and table values are reproduced from the cited articles for comparison and are © their authors and publishers. The upstream reference implementation and demo dataset are credited to the DSI authors.

1. Moran PAP (1950). Notes on continuous stochastic phenomena. Biometrika 37(1–2):17–23. doi:10.1093/biomet/37.1-2.17
2. Wang J-F, Li X-H, Christakos G, et al. (2010). Geographical detectors-based health risk assessment… IJGIS 24(1):107–127. doi:10.1080/13658810802443457
3. Anselin L (1988). Spatial Econometrics: Methods and Models. Springer. doi:10.1007/978-94-015-7799-1
4. Zhang Z, Song Y, Luo P, Wu P (2023). Geocomplexity explains spatial errors. IJGIS 37(7):1449–1469. doi:10.1080/13658816.2023.2203212 · PDF
5. Liu H, Song Y, Yi W (2026). Degree of spatial interpretability. IJGIS. doi:10.1080/13658816.2026.2614335 · PDF (source of Table 3 and the Fig 1/Fig 6 reference images on this page)
6. Liu H, Song Y, Yi W, Zhang P (2026). Degree of geocomplexity for diagnosing spatial pattern loss beyond prediction accuracy. GIScience & Remote Sensing 63(1):2657087. doi:10.1080/15481603.2026.2657087 · PDF (source of the DG Fig 8 reference image on this page)
7. Reference implementation and demo dataset: github.com/KevinOceanLiu/DSI