0.14.2 - Licensing metadata release. Adopted split licensing: the
reusable core/PyPI package remains MIT, while the QGIS plugin
integration and plugin distribution are GPLv3.
0.14.1 - Robustness fixes (no API additions). The stochastic precision
stop criterion now uses the Wilson 95% confidence-interval error on
the reported fraction - max(p_hat - lower, upper - p_hat) - instead of
a standard error, which reduces (to the chosen confidence level, not
with certainty) the chance of a rarely-sampled route triggering
premature convergence and a falsely deterministic map; repeated
interim checks are not a formal confidence sequence, so effective
coverage is somewhat looser than the nominal 95% (note: tol is now a
CI error bound, stricter than the old standard error). Convergence
parameters are validated (tol finite & > 0; min_iter / check_every /
patience integer >= 1; check_every=0 no longer divides by zero,
patience=0 no longer false-converges). Circuit current density returns
a null map when a source is also a grounded target (was near-uniform
phantom current). Multi-criteria out_range now requires 0 < lo < hi,
finite (a 0 minimum made cells impassable; a negative one broke the
geometric mean). Non-finite (NaN / inf) cost-function weights are
rejected with a clear message instead of surfacing as NumPy
'Probabilities contain NaN'. The Llobera & Sluckin cost function no
longer applies abs() to its linear term: the published quartic
(Llobera & Sluckin 2007, Eq. 16) uses a signed slope, so the cost
minimum sits on a gentle downhill (~ -0.175) instead of being forced
to flat ground - this restores the downhill anisotropy and lowers
descent costs (uphill unchanged). Scientific-correctness audit fixes:
FETE now computes all n*(n-1) directed routes (was unordered pairs,
dropping anisotropic return legs); RSP makes the target absorbing per
the cited Saerens/Panzacchi/gdistance process (was a shared
non-absorbing matrix, which could reorder passage values); PDI now
divides by the straight-line origin-destination distance, snaps the
modelled path's endpoints to the reference O/D first (per
leastcostpath), and is attributed to Jan, Horowitz & Peng 2000 (was
reference-path length, no snapping, Goodchild & Hunter). Docs
corrected: the wheeled/pack-animal presets are labelled experimental
Itinera heuristics (Herzog's vehicle function is symmetric; no
validated pack-animal function exists), the stochastic DEM error is an
Itinera variogram-FFT field in the spirit of (not a reproduction of)
Hunter & Goodchild / Lewis, the multi-criteria builder is a generic
Itinera heuristic with an extent-dependence caveat, and the precision
criterion is a pointwise (not simultaneous) 95% interval.
0.14.0 - Movement-mode cost presets + accessibility surfaces
(completes the methods roadmap). Two anisotropic critical-slope cost
functions - Wheeled (cart, ~8% critical upward slope) and Pack animal
(~25%) - where cost rises quadratically with grade and the uphill
limit is tighter than downhill (Herzog 2013, Verhagen 2019); ten cost
functions total. A new Accessibility / cost catchment algorithm
computes the cost-distance surface from source point(s) plus an
optional within-budget catchment mask and isochrone bands
(core/accessibility.py).
0.13.0 - Convergence / stop criterion + progress reporting for the
stochastic LCP. The Monte-Carlo loop can stop early once the
probabilistic corridor is good enough: stabilisation (the map stops
changing, max|dp| < tol) or precision (max binomial standard error <
tol). The Stochastic LCP algorithm gains Maximum iterations +
convergence tolerance + criterion + minimum iterations, and logs the
live metric per checkpoint. (stochastic_lcp gains tol / convergence /
return_diagnostics.)
0.12.0 - Cost-model randomisation in the stochastic LCP (Herzog 2022:
no universal best cost model). The Stochastic LCP cost function is now
multi-select - each Monte-Carlo realisation samples one cost function
from the chosen set (uniform or weighted) - plus a parameter-jitter
control that perturbs the Pandolf mass/load/terrain by a fraction each
iteration. The probabilistic corridor now integrates cost-model
uncertainty alongside DEM error and edge-dropping (stochastic_lcp
gains cost_fns / cost_weights / param_jitter).
0.11.0 - Variogram-based DEM error fields (Hunter & Goodchild 1997).
The stochastic DEM-error model is now a true geostatistical Gaussian
random field - exponential (new default), spherical or gaussian
variogram with an optional nugget - generated by FFT spectral
simulation (scipy.fft, no gstat dependency), replacing the
Gaussian-filter approximation (kept as a fast option). The Stochastic
LCP algorithm gains DEM error model + nugget parameters, and a new
standalone DEM error realisation tool outputs one perturbed DEM
(core/stochastic.py::simulate_error_field).
0.10.0 - Multi-criteria composite friction (Herzog 2022, Litvine
2024): a new Cost-surfaces algorithm and core/multicriteria.py that
merge several penalty rasters (hydrology, wetness, land cover,
viewshed masks) into one friction multiplier - each min-max
normalised, optionally inverted and weighted, combined by a weighted
arithmetic (sum) or geometric (product) mean into a chosen range (1
neutral, >1 discourages, <1 prefers); NoData is impassable. Plugs into
the existing multiplier / friction slots. Completes Tier 2 of the
methods roadmap.
0.9.0 - Circuit-theory connectivity (McRae 2008, 2012): movement as
electrical current flow, the random-walk complement to the LCP. Two
algorithms in a new Connectivity group plus core/circuit.py (pure
numpy/scipy): Circuit current density / pinch points (solve the graph
Laplacian Lv=i, source injected / target grounded, for a
current-density raster + optional pinch points within the least-cost
corridor; the anisotropic conductance is symmetrised, use RSP for the
directed current) and Connectivity barriers / restoration (a
moving-window improvement map over the corridor surfaces marking the
strongest barriers).
0.8.0 - Randomized Shortest Paths (RSP): a single theta parameter
spans the whole optimal-to-random movement axis (large theta =
least-cost path, small theta = random-walk / circuit current density;
Panzacchi 2015, van Etten 2017). New core/rsp.py (pure numpy/scipy: W
= P_ref * exp(-theta*C), one sparse LU of (I-W) over the existing
conductance matrix) and a Processing algorithm (one origin to
destination(s); a movement-density raster + the RSP free-energy
distance; a 'Normalise costs' flag so theta ~ 1 is meaningful across
cost functions). Keeps anisotropy throughout.
0.7.1 - Hardening + docs. Buffer validation now rejects non-positive
buffer distances (a 0 distance could previously make the densifier
allocate a near-unbounded number of points on long lines); the user
docs (README, PyPI description, manual) are brought up to date with
the 0.7.0 features (eight cost functions, Buffer Validation,
Sensitivity Analysis).
0.7.0 - Three method additions completing the review's Tier-1 roadmap.
(1) Energetics cost functions: Irmischer & Clarke (GPS-calibrated
walking speed), Minetti (cost of transport) and load-aware Pandolf
with the Santee/Yokota downhill correction; cost functions now accept
extra keyword parameters (body mass / load / terrain factor), exposed
on every conductance-building algorithm. (2) Buffer-overlap validation
(Goodchild & Hunter 1997) as a multi-distance similarity table, beside
the PDI. (3) Sensitivity analysis: sweeps the selected cost functions
x connectivities for one origin/destination pair and reports an
agreement raster, a per-configuration summary table, optional
individual paths, and a route-stability scalar.
0.6.1 - FETE can now optionally output the individual least-cost paths
as a line layer (one feature per point pair, with from_id/to_id/cost),
alongside the traversal-frequency raster. The paths were already
computed internally; they are now exposed via an optional output. No
change to the raster result.
0.6.0 - The pure numpy/scipy core is now also published as the
`itinera` PyPI library (pip install itinera), built single-source from
core/ via hatchling. No change to the QGIS plugin itself.
0.5.9 - Dev tooling: flake8 is now a CI check (max-line-length 88);
excluded maintainer/dev-only files (CLAUDE.md, setup.cfg, pytest.ini,
requirements-dev.txt) from the packaged zip. No runtime change.
0.5.8 - QGIS 4 / Qt6 runtime fixes: the settings dialog crashed on
QDialogButtonBox.Ok (now scoped StandardButton), and the LCP output
used QVariant.Int which PyQt6 removed (now QMetaType on Qt6, QVariant
on Qt5). Verified on QGIS 3.28 and 4.0.
0.5.7 - Code hygiene: removed unused imports (flake8 F401) and added a
setup.cfg ignoring W503/W504. No behaviour change.
0.5.6 - Declared QGIS 4 compatibility via qgisMaximumVersion=4.99
(without it QGIS assumed a 3.99 maximum and marked the plugin
incompatible on QGIS 4). Verified loading on QGIS 4.0.
0.5.5 - Documentation: added QGIS 3.28+/4.0 and Qt5/Qt6 compatibility
badges to the README.
0.5.4 - QGIS 4 / Qt6 compatibility: the settings dialog used exec_(),
which PyQt6 removed; switched to exec() (works on QGIS 3 and 4). No
other changes.
0.5.3 - The "Interactive LCP settings…" button now has its own gear
icon, distinct from the path icon of the LCP tool button.
0.5.2 - Replaced the blank placeholder plugin icon with a real,
visible one, so the two interactive-tool buttons are findable on the
Plugins toolbar. Clarified in the docs where those buttons live.
0.5.1 - Documentation only: README status badges, future-directions
wording, and removal of the internal publishing guide from the README
and the packaged zip. No code changes.
0.5.0 - New "Resample DEM (block mean)" algorithm and a memory
pre-flight warning for large DEMs; stochastic_lcp guards against
n_iter < 1; added a user manual (docs/MANUAL.md) and verified
references (docs/REFERENCES.md, references.bib). Corrected stochastic
citation to Lewis 2021.
0.4.0 - Interactive LCP tool is now configurable: an "Interactive LCP
settings…" toolbar/menu action picks the cost function and
neighbourhood (parity with the Processing algorithms); the graph cache
rebuilds only when settings change.
0.3.0 - New Stochastic least-cost path (Lewis 2023): N Monte-Carlo
realisations with spatially-correlated DEM error (RMSE-scaled) and/or
random edge dropping, accumulating a probabilistic corridor in [0,1].
Seed for reproducibility; supports the barrier/multiplier raster.
0.2.2 - Fix: xy_to_rowcol now floors instead of truncating, so points
just west/north of the raster are correctly out of bounds (not wrongly
snapped to row/col 0).
0.2.1 - Fixes: point layers are reprojected to the DEM CRS before cell
lookup (algorithms + map tool); barrier/friction rasters are validated
against the DEM CRS and geotransform (not just pixel count);
rotated/non-square rasters are rejected with a clear error. Documented
PDI limitations.
0.2.0 - Optional barrier / multiplier raster on the slope-based
algorithms (slope cost surface, LCP, LCC, FETE): edge cost is scaled
by the mean of the two cells' values (>1 discourages, <1 prefers, e.g.
known roads); NoData or <=0 cells are impassable (cliffs, deep wadis).
0.1.0 - Initial release: anisotropic LCP, LCC corridor, FETE, PDI
validation, slope & friction cost surfaces, interactive two-click LCP
map tool. Pure numpy/scipy/GDAL (no external pip dependencies).

None