[general]
name=02SpatialML - Spatial Machine Learning Lab
qgisMinimumVersion=3.28
qgisMaximumVersion=4.99
description=Predictive mapping with machine learning in QGIS: spatial cross-validation designs with distance diagnostics, five models on the same folds, explanations on held-out data, a wall-to-wall prediction with interval and area of applicability, and a replayable model record.
version=0.1.1
author=Geophilo
email=yusuf.eminoglu@deu.edu.tr
about=02SpatialML carries one predictive-mapping study from samples to a defensible map in one dock panel, with no extra installs.
    Data: point or polygon samples with a target field, covariates from the attribute table and from a stack of covariate rasters harmonised onto one grid; distances in metres in a projected working CRS.
    Validation design: random k-fold, square or hexagonal spatial blocks, coordinate clusters, buffered leave-one-out and kNNDM, with block size and buffer radius taken from the variogram range, and nearest-neighbour distance curves (sample, prediction and CV) with the Wasserstein distance that show whether the folds mimic the prediction task.
    Models: OLS baseline, random forest with quantile regression forest intervals, gradient boosting with split-conformal intervals, GWR and MGWR, tuned by nested cross-validation on the same folds; leaderboard with the optimism gap between spatial and random cross-validation and the residual Moran's I.
    Explanations: permutation importance on held-out folds, exact TreeSHAP, partial dependence with ICE curves, accumulated local effects and local GWR / MGWR coefficients.
    Prediction: raster or layer prediction with interval bounds and width, the dissimilarity index and the area of applicability mask (Meyer and Pebesma 2021), styled in a layer group.
    Record: a self-contained HTML model card with KaTeX formulas and a JSON run manifest that replays the study and checks every hash and metric.
    Every engine is implemented in numpy; SciPy and scikit-learn are used when present. Hot spots, LISA, kriging and other learners stay in GeoStats Lab.
    Manual: https://geophilo.com/zero2spatialml/
repository=https://gitlab.com/geophilo1/zero2spatialml
tracker=https://gitlab.com/geophilo1/zero2spatialml/-/work_items
homepage=https://geophilo.com/zero2spatialml/
category=Raster
tags=machine learning,spatial machine learning,predictive mapping,spatial cross-validation,cross validation,area of applicability,dissimilarity index,random forest,quantile regression forest,gradient boosting,gwr,mgwr,geographically weighted regression,shap,partial dependence,accumulated local effects,permutation importance,variogram,knndm,model card,reproducibility,raster,geophilo
icon=icons/icon.png
experimental=False
deprecated=False
hasProcessingProvider=yes
license=GPL-2.0-or-later
changelog=
    0.1.1 - MGWR minimum bandwidth (dock, Processing, run record): the demo sets 100 neighbours so MGWR back-fitting converges on its clustered samples; the model card states the bound and whether back-fitting converged. Kit 1.2: chart legends in their own band above the plot, transparent rail buttons under the Night Mapping and Blend of Gray themes, dock tests under the real QGIS themes. Blend of Gray: readable list text and no grey band behind disabled fields; Target KPI shows the mean with the range below.
    0.1.0 - First release. Sample table from point or polygon samples with attribute and raster covariates, raster harmonisation, variogram, six validation designs with nearest-neighbour distance diagnostics, OLS, random forest, gradient boosting, GWR and MGWR on shared folds with nested tuning, permutation importance, TreeSHAP, PDP / ICE and ALE, raster and layer prediction with intervals, dissimilarity index and area of applicability, HTML model card, run manifest with replay, demo landscape.
