Skip to content

Releases

HABITUS v2.1.0

Choose a tag to compare

@omerorucu omerorucu released this 13 Nov 22:05

v2.1.0 — 14 November 2026

HABITUS is now free software under the GNU General Public License, version 3.
The Linux packages no longer close on a key press. The ensemble map contains
the models that the run record says were projected. ⑥ Range change can count
change within the accessible area, with or without cells in extrapolated
conditions, and always reports how much of the change lies in such cells.
Tables separated by semicolons or tabs are read. Results change for ensemble
maps (see below); everything else gives the same numbers as 2.0.2.

Ensemble map: the projected models only (results change)

Cause. The projector built the EM* maps from every model that passed the
quality filter, the cross-validation (fold) models included, while the
per-algorithm maps, the "Also project the cross-validation models" option
(off by default) and the run record ("replicates_projected": "full") said
that only the models refitted on all data were projected. When an
algorithm's full model fell below the minimum score while some of its fold
models passed, the algorithm entered the ensemble through those fold models
alone. Found while reproducing a published biomod2 ensemble.

Now. The ensemble combines the projected models: the full models, plus
the fold models only when the option is ticked. An algorithm whose full model
does not reach the minimum score is left out; when no full model reaches it,
no ensemble map is written and a notice names the option. Ensemble maps
differ slightly from 2.0.2 (fewer members, no fold models); cross-validated
ensemble scores are unchanged.

Range change within the accessible area

Cause. ⑥ compared the current and future maps over the whole study
boundary. With a boundary drawn around several species, or much larger than
the area a species could reach, most of the projected loss lay far from any
record and in conditions outside the calibration range. In a case study of
four firs, the net change of one species under SSP5-8.5 was −55% over the
boundary and −16% within 300 km of its records, where 98% of the cells were
extrapolated.

Now. ① Data writes the accessible area as accessible_area.tif. ⑥ reports,
besides the whole boundary, the change within the accessible area and within
it without cells where ExDet NT1 < 0, and the share of changed cells in
extrapolated conditions; "Count change within" sets the extent of the
headline numbers and of the map. The default (whole boundary) and its
numbers are unchanged. The report lists the same, with a note when a quarter
or more of the change is extrapolation.

Linux: crash on a key press

Cause. The Linux packages carried libraries of the build system (Ubuntu
22.04) that the user's system loads its own companions for. The bundled
libxkbcommon, paired with the system's newer libxkbcommon-x11, crashed
HABITUS on Shift, Ctrl or any key on newer distributions; the bundled GLib
made the system's gvfs module fail ("undefined symbol:
g_task_set_static_name").

Now. libxkbcommon, the X11/xcb core libraries, GLib/GIO, D-Bus,
fontconfig, FreeType, HarfBuzz, libstdc++ and the GL libraries are taken from
the system (every supported system, glibc 2.35 or later, has them).

Turkish and other locales: GDAL errors

Cause. Qt sets the user's locale for the whole process on Linux and
macOS. In a Turkish locale, GDAL's case-insensitive comparisons turn "I" into
a dotless "ı": hundreds of "apszKnownConfigOptions[] isn't correctly sorted"
errors were printed, and names differing only in case could stop matching.

Now. Character handling, collation and numbers return to the neutral C
locale after the interface starts; the interface itself is unaffected.

Crash log

A native crash or an uncaught error is written to crash_YYYYMMDD.log under
the user data folder (Windows %LOCALAPPDATA%\HABITUS\logs, macOS
~/Library/Application Support/HABITUS/logs, Linux
~/.local/share/HABITUS/logs), so that a program that closed without a
message leaves something to send.

Tables separated by semicolons or tabs

Spreadsheets in many European locales save "CSV" with ";" between fields and
"," as the decimal mark, and GBIF downloads are tab-separated. Every
occurrence, absence and target-group table is now read with the delimiter and
decimal mark of the file, and a UTF-8 byte-order mark is removed. GBIF
columns (gbifID, species, decimalLongitude, decimalLatitude ...) are found by
name, and the species name is no longer taken from a numeric identifier
column.

Messages

  • A batch run's summary and log keep the reason a species failed (e.g. "The
    29 presence record(s) fall in only 3 predictor cell(s) …"); before, they
    kept the first lines of the error trace and lost the reason.
  • A predictor layer that has no value at any record is named; before, the
    message suggested checking the coordinate system of the records.
  • With a single predictor, ② Variables says that modelling needs two;
    before, it stopped with an error trace and Confirm refused.
  • A warning that the program logs itself is no longer written to the log a
    second time.
  • The log line of a full model no longer prints a one-split score that the
    results replace by the mean of the cross-validation folds.
  • ⑥ Fill Paths says which list to choose instead of doing nothing.
  • The run record labels values without a shipped default (inputs and numbers
    found during the run) "recorded" instead of "user", and records one record
    per cell with its default.

Licence

GNU General Public License, version 3 (previously MIT). The corresponding
source code is provided on request (omerorucu@sdu.edu.tr).

The macOS packages of 2.1.0 follow later; until then macOS users keep 2.0.2.

HABITUS v2.0.2

Choose a tag to compare

@omerorucu omerorucu released this 10 Nov 18:32

v2.0.2 — 10 November 2026

Occurrence records are reduced to one per predictor cell by default, and
spatial thinning gains two presets. Variable recommendation, spatial block
folds, the fewest presences a model is fitted from, and LASSO/Ridge in
Advanced Analysis are revised. A step that cannot be carried out now stops
the run or says what was done instead; before, several were replaced by a
different step without notice. Whole-world predictors are read within the
memory of an ordinary computer, and the Linux packages start the program
correctly. Results change for any occurrence file in which two or more
records fall in the same predictor cell.

One record per raster cell (new default)

Cause. ① Data removed only records with exactly the same coordinates.
Records a few metres apart, which is how a much-visited site appears in GBIF
and other aggregated data, all stayed in, although every one of them falls in
the same predictor cell and carries the same predictor values. Each extra
record is a copy of the first: it adds weight to that cell without adding
information. The message "no duplicates" made it look as if the records had
been checked for this. Found on GBIF records of a Mediterranean fir: 628
records fell in 31 cells, one of them holding 226; one cross-validation fold
received 606 of the presences and the others 7 or 8.

What was changed.

  • ① Data has a new option, Keep one record per raster cell, on by default.
    The first record of each cell, in file order, is kept; the others are listed
    among the excluded records with the reason "same raster cell as another
    record", shown on the Map tab, and counted in the log, the report, the run
    record (points_dropped_by_reason, one_record_per_cell) and the ODMAP
    export. It is applied before thinning.
  • With the option off, the number of records that share a cell is stated in
    the processing notices.
  • An external test set is evaluated as supplied; the option applies to the
    training records only.

Thinning distance

  • Fixed distances. Besides a distance typed by the user and the automatic
    distance from the variogram, ① Data offers One cell (the predictor
    resolution at the records' mean latitude) and 10 km, the distance used by
    Boria et al. (2014, doi:10.1016/j.ecolmodel.2013.12.012). Neither is a rule
    for every species; the distance used and how it was chosen are recorded.
  • A typed distance smaller than one predictor cell is pointed out.
  • When the automatic distance would remove more than half the records and is
    skipped, the message now says whether records sharing a cell were reduced.

Recommended variables

Cause. "Recommended" in ② Variables ticked the variables whose VIF in the
full set was at most 10. With the 19 bioclimatic layers nearly every VIF in
the full set is far above 10, because each layer is close to a combination of
the others, so only a few were ticked: on one data set three temperature
variables and no precipitation.

What was changed. "Recommended" is now a stepwise selection: the variable
with the highest VIF is removed and VIF recomputed until every VIF is at or
below the threshold (as usdm::vifstep; Naimi et al. 2014), then the higher-VIF
member of any pair above the correlation threshold is removed. The removals
are written to the log in order, and variable_priority_ranking.csv gains
the column VIF_in_selection. Variables with no strong pairwise correlation
are no longer added on top. The selection is still only a starting point.

Spatial block folds: absences balanced as well

Cause. Whole blocks were assigned to folds so as to even out the number
of presences only. Every block without a presence added nothing to a fold's
load, so all of them went to the same fold. With presences in a small core
and pseudo-absences spread over a wide buffer, one fold received almost all
the pseudo-absences: 741 against 23 on one data set, 978 of 1000 in a test
case. Per-fold prevalence then differed several-fold, and TSS, the threshold
and the ensemble's minimum-score filter with it.

What was changed. Blocks holding presences go first, each to the fold
whose share of presences (counted twice) plus share of absences is lowest after
adding it; blocks without a presence then go to the fold with the fewest
absences, so the fold that holds a densely recorded block still receives its
share of absences. 50 seeded orderings are tried and the most even kept, as
blockCV does with its iterations. The CV note in the log and the report now
gives the presences and absences per fold. Checkerboard, band, environmental
and random folds are unchanged.

Fewest presences: 5

Cause. A species with three usable presences was modelled, with a
warning. Spatial blocks cannot be formed from three records, so the folds
were random, each scored on a single presence; AUC values of 0.96 and 1.00
came out of two such species in a test run.

What was changed. At least 5 usable presences are now required (3
before), the smallest sample in the published evaluations of SDM methods at
small sample sizes (Hernandez et al. 2006, doi:10.1111/j.0906-7590.2006.04700.x;
Pearson et al. 2007, doi:10.1111/j.1365-2699.2006.01594.x). With fewer than 25
presences the processing notices recommend jackknife evaluation (leave one
presence out), which Pearson et al. (2007) proposed for such samples. When
nothing was lost to the raster extent or NoData, the message no longer
suggests checking the coordinate system. An external test set still needs 3.

Advanced Analysis: LASSO and Ridge as logistic regressions

Cause. In ② Variables > Advanced Analysis, LASSO and Ridge were
least-squares fits to the 0/1 presence label, not models of a binary
response. The Ridge "recommended" list was the variables above the median
coefficient, which has no basis. Pooling the PA repetitions counted each
presence once per repetition, and in the cross-validation of the penalty a
copy of a presence could sit in the training and the test fold at once.

What was changed.

  • Both are penalised logistic regressions (L1 and L2) of presence against
    absence on the standardised predictors, with presences and absences
    weighted to equal totals (Barbet-Massin et al. 2012). The penalty is on the
    glmnet scale and chosen by stratified k-fold cross-validation on binomial
    deviance, as cv.glmnet does.
  • LASSO selects at the 1-SE rule by default (the largest penalty within one
    standard error of the minimum deviance); the minimum or a typed value can
    be chosen instead. Ridge, which removes no variable, ranks them and no
    longer recommends any.
  • The plots show the cross-validated deviance under the coefficient paths;
    Save All writes the full path (lasso_path.csv, ridge_path.csv) and the
    penalties to lasso_info.txt.
  • Each presence is counted once when the PA sets are pooled (also for the
    condition number and PCA).
  • Checked against R glmnet 5.1 (family = "binomial") on simulated data:
    coefficients within 3e-3 (LASSO) and 4e-6 (Ridge) over the path, the same
    variables set to zero.

Smaller changes

  • The "records per predictor" notice in ① Data now says that it counts the
    predictors loaded, and how many the records allow, since the number kept in
    ② Variables is what matters.
  • The records-per-predictor check is repeated with the predictors actually
    kept, when a selection is confirmed in ② Variables and when the multi-species
    batch run applies the first species' selection to another species. Before,
    a species with 22 records modelled on 8 predictors chosen for another
    species got no notice. The notice is written to the log and the report.
  • Multi-species batch run: each species now gets its own session log in its
    own folder, and its report its own Session Log section. Before, the log
    opened for the first species stayed open, every later species was written
    into it, and their reports had no session log.
  • Maps on screen: the legend of the occurrence points (training presences,
    test presences, pseudo-absences) is drawn below the map instead of in its
    upper-right corner, where it covered the points it describes. The colour
    bar takes the height of the map; with a wide, low extent it used to run far
    above and below it.
  • Variables > Advanced Analysis: Save All Results wrote an incomplete
    lasso_info.txt (the list of selected variables was lost) and reported a
    warning; pca_loadings.csv carried an empty column. Both fixed.
  • Random folds (requested, or used when spatial blocks cannot be formed) are
    never more than the presences or absences: with three presences and four
    folds, one fold used to hold no presence, could not be scored, and the
    others were scored on single presences. The note says when fewer folds
    than requested were made.
  • A data set whose records fall in fewer than three predictor cells is now
    stopped with a message saying so, instead of a hint about coordinate
    systems.

Results that change. Any run whose occurrence file has more than one record
in a cell: fewer presences, different pseudo-absence and background samples,
different folds and scores. Files with at most one record per cell give the
same results as v2.0.1. To repeat a v2.0.1 run, switch the option off.
Runs whose variables were chosen with "Recommended" can get a different set.
Every run with spatial block cross-validation (the default) gets different
folds, and so different cross-validation scores, even with the same seed.
Species with 3 or 4 usable presences are no longer modelled. Advanced
Analysis gives different LASSO and Ridge coefficients and selections; it
does not affect the models unless a variable set was chosen from it.

Accessible area smaller than the number of points requested

Cause. When the accessible area (buffer, buffered minimum convex polygon
or calibration polygons) held fewer usable cells than the background or
pseudo-absence points requested, the points were d...

Read more