GeoStats Lab finds hot spots and cold spots (Getis-Ord Gi*), clusters and spatial outliers (local Moran's I, LISA), tests whether a pattern is spatially autocorrelated (global Moran's I, Geary's C, General G), groups areas into contiguous regions (SKATER, max-p, AZP, REDCAP), fits spatial regression models and interpolates surfaces, all as Processing tools: 102 of them in eight groups, ordered the way a study runs. Results are written as layer fields with documented meanings, and many tools add an HTML report that explains the numbers in plain language. In the hot spot and cluster maps, features with a missing value or with no neighbours are shown as their own class and counted in the report instead of being hidden or called not significant; contiguity weights on points and a distance band on a geographic CRS (degrees) are refused with a hint. Global Moran's I, General G, Geary's C, Gi*, local Moran's I and the contiguity weights are checked against PySAL esda in the test suite.
00 SETUP AND DIAGNOSTICS (5): Is the data fit for spatial statistics, and which tool answers my question? A data readiness audit, a workflow advisor that suggests which tools fit your question and data, a sample dataset guide, and two library tools: GeoStats Library Status checks what is installed without changing anything, and Install / Update GeoStats Libraries installs optional packages only when you run it.
01 DATA PREPARATION AND NEIGHBORHOODS (3): Every spatial statistic depends on who counts as a neighbour, so this group makes that choice explicit and testable. A spatial weights workbench with distance-band selection, empirical Bayes rate smoothing for small-population areas, and attribute export to CSV or ASCII.
02 URBAN PATTERN SCAN (12): Is the pattern clustered, dispersed or random, at what distance, and does it hold up under a different zoning? Global Moran's I, incremental spatial autocorrelation to find the distance at which clustering peaks, Geary's C, Getis-Ord General G, join count statistics, the gamma index, global bivariate Lee's L, average nearest neighbour, Ripley's K, the Geodetector q-statistic for stratified heterogeneity, Gini and spatial Gini, and S-maup, which tests whether your result depends on the zoning you happened to use.
03 HOT SPOTS, SPATIAL OUTLIERS AND REGIONS (18): Where exactly are the clusters and the places that break the pattern, and how can areas be grouped into coherent regions? Getis-Ord Gi* hot spots; local Moran's I and local Geary's C clusters and outliers; bivariate LISA and Lee's L; empirical Bayes LISA for rates; local join counts for binary clusters and co-location; partial and conditional local Moran that removes the effect of covariates; local spatial heteroskedasticity (LOSH); multivariate local Geary; the colocation quotient; similarity search; k-means multivariate clustering; and regionalization with SKATER, AZP, REDCAP and max-p regions that meet a minimum size, with a quality assessment for the zones they produce.
04 CENTERS, DIRECTION AND DISPERSION (6): Where is a distribution centred, which way does it stretch, and how far does it spread? Mean centre, median centre and central feature, standard distance, the standard deviational ellipse and linear directional mean.
05 MODELS AND SCENARIOS (15): What explains the pattern, does the relationship change from place to place, and how sure can we be? OLS, generalized linear regression, Lagrange multiplier tests that tell you which spatial model the data call for, spatial lag, spatial error and spatial Durbin models, spatial regime and quantile regression, exploratory regression across variable combinations, eigenvector spatial filtering, geographically weighted summary statistics, GWR and multiscale GWR (MGWR), a model comparison matrix, and an attribute randomization test that shows whether a pattern survives shuffled data.
06 MACHINE LEARNING AND EXPLAINABLE AI (34): When relationships are non-linear, these models predict and classify, and the explainability tools show why each prediction came out as it did, so a model can be discussed in a planning review rather than taken on trust. Regression and classification with random forest, extra trees, gradient boosting (scikit-learn, XGBoost, LightGBM and CatBoost), support vector machines, neural networks and the TabPFN tabular foundation model, plus DBSCAN, HDBSCAN and Gaussian mixture clustering. Models are validated with spatial k-fold or kNNDM cross-validation and compared on a leaderboard, and their residuals are checked for leftover spatial autocorrelation. Explanations come from SHAP (global importance, a spatial attribution map and local reports), permutation importance, partial dependence, glass-box Explainable Boosting Machines and DiCE counterfactuals, and every prediction can carry a distribution-free conformal interval and an uncertainty map.
07 INTERPOLATION AND GEOSTATISTICS (9): How do you turn sample points into a continuous surface and know how far to trust each cell? Empirical semivariogram and model fitting; ordinary and universal kriging that writes its own variance raster; empirical Bayesian kriging, which accounts for the uncertainty of fitting the variogram; indicator kriging for the probability of exceeding a threshold; IDW, natural neighbour, radial basis functions and polynomial trend surfaces; and a comparison tool that scores the interpolators under one shared validation protocol.
The statistics, models and geostatistics run on the Python that ships with QGIS. The machine-learning tools use optional libraries such as scikit-learn, XGBoost, LightGBM, CatBoost, SHAP and TabPFN; they are installed only when you run Install / Update GeoStats Libraries, and GeoStats Library Status shows which ones are present.
Manual: https://geophilo.com/planx_geostats/manual.html