② Parameters tab¶
The Parameters tab controls the modeling and evaluation settings: which Maxent feature classes to allow, the regularization multiplier, the spatial cross-validation method, and where to write outputs. Defaults follow established Maxent practice; this chapter explains what each control does and when to override the default.
Step by step¶
Step 1. Check the main settings¶

- Feature Types: keep Auto (maxnet default rule by sample size) unless you have a reason to choose the feature classes yourself.
- Regularization: the Regularization multiplier controls how smooth the model is. Larger values give simpler models.
- Spatial evaluation: the Method used to test the model, the number of Folds, and Fix random seed for repeatable results.
Step 2. (Optional) Choose the feature classes yourself¶

- Select Manual selection.
- Tick the feature classes to use: Linear, Quadratic, Hinge, Product and Threshold.
Step 3. (Optional) Choose another evaluation method¶

- Open the Method list. Geographic K-Fold is the default and the safest choice for most data. The other methods are described in Spatial cross-validation.
Step 4. Check the importance options and the output files¶

- Advanced: hinge and threshold knots, Add presences to background and Down-weight spatially clustered points. The defaults suit most data.
- Fix random seed keeps the folds and the background sample the same from run to run.
- Jackknife variable importance fits extra models to measure each variable. It makes the run longer.
- Permutation importance measures how much the model gets worse when each variable is shuffled.
- Output Files: where the model (
.pkl) and the results workbook (.xlsx) are saved. Click ... to change the paths.
Then click ▶ Run Maxent at the bottom of the dock.
Feature Types¶
Maxent's "features" are the basis functions it uses to express the response of suitability to each environmental variable (Phillips, Anderson & Schapire 2006; Phillips et al. 2017). The five classes (LQPHT) are:
| Feature | Symbol | What it captures |
|---|---|---|
| Linear | L |
Monotonic linear response |
| Quadratic | Q |
Optimum-shaped (∩) or U-shaped (∪) response |
| Product | P |
Pairwise interactions between two variables |
| Hinge | H |
Piecewise-linear "kink" — sharp changes at threshold values |
| Threshold | T |
Step-function — a hard cut-off at some value |
QMaxent offers two modes:
-
Auto (maxnet default rule by sample size) (default): uses the default feature classes of maxnet, the R implementation of Maxent by its authors (Phillips et al. 2017;
maxnet.formula(classes = "default")), based on the number of presence points:Presences Feature classes fewer than 10 L 10 to 14 LQ 15 to 79 LQH 80 or more LQPH Threshold features are not added automatically. - Manual selection: tick any subset of LQPHT. The boxes start with L, Q, P and H ticked and Threshold unticked. Use this to compare alternatives or to reproduce a published study that fixed a particular feature combination (e.g. the Pitta nympha example fixes feature classes to LQH to match Lee et al. 2025).
Why Auto no longer adds threshold features
Up to v0.1.x, Auto followed the maxent.jar rule and added threshold features (LQPHT) for 80 or more presences. With elapid's scikit-learn backend, every feature gets the same regularization strength, while maxent.jar regularizes each feature class separately. Threshold features then overfit. On the Bradypus benchmark (random 10-fold CV), the test AUC was 0.873 with LQPHT and 0.890 with LQPH; maxent.jar with LQPHT reached 0.895. v0.2.0 therefore follows the maxnet default, which does not use threshold features. They remain available under Manual selection.
Auto is a default, not the best model
The rule gives a reasonable starting point, but the most suitable feature classes depend on the data (Merow, Smith & Silander 2013; Radosavljevic & Anderson 2014). To choose them, compare alternatives with Manual selection and spatial cross-validation, or use a model-selection tool such as ENMeval (Muscarella et al. 2014). Be aware that repeatedly tuning settings against random (non-spatial) CV gives optimistic AUCs (Roberts et al. 2017).
Regularization multiplier¶
Maxent fits a regularized maximum-entropy distribution — the regularization multiplier (RM) controls how strongly the fit is penalised for complexity. - RM = 1 (default) — the value Phillips & Dudík 2008 showed yields the best held-out AUC across the full Maxent benchmark. - RM > 1 (e.g. 2–4) — smoother responses, less over-fit, lower training AUC but often higher CV AUC. Use when occurrences are small in number, spatially clustered, or biased. - RM < 1 — only justified by a formal hyperparameter search such as ENMeval (Muscarella et al. 2014) or Wallace (Kass et al. 2018).
Spatial cross-validation¶
A drop-down picks the spatial CV scheme used to split presences into folds:
| Method | Best for | Reference |
|---|---|---|
| Geographic K-Fold (default) | General-purpose | Roberts et al. 2017 |
| Random K-Fold | Reproducing studies that used random splits | Phillips et al. 2017 |
| Checkerboard | Strong spatial autocorrelation in presences | Muscarella et al. 2014 |
| Buffered Leave-One-Out | Very small datasets (n < 30) | Hijmans 2012; Valavi et al. 2019 |

The default Geographic K-Fold (k = 5, fixed seed = 42) strikes a balance recommended by Roberts et al. 2017: geographic blocks break the spatial-autocorrelation cheating that random K-Fold permits, yet keep the held-out folds large enough for a stable AUC.
The fixed seed makes every run bit-for-bit reproducible — a property cited as essential by the SDM-best-practice review of Araújo et al. 2019.
Jackknife variable importance¶
When the Jackknife checkbox is on, QMaxent fits 2 × N additional single-predictor and leave-one-out models to estimate each variable's unique contribution. The cost is roughly 2 × N × training time; for the Bradypus dataset (9 variables, 116 presences) this adds about 30 seconds.
Jackknife output is rendered in two places: 1. The ④ Results → Jackknife Importance sub-tab as a paired-bar plot. 2. Table 4 of the exported XLSX workbook, with all four AUC columns plus computed Train AUC drop and Test AUC drop.
The interpretation follows the original Phillips, Anderson & Schapire 2006 reading: "with-only" bars show univariate signal; "without" bars show how dependent the full model is on each variable.
Permutation importance¶
When Permutation importance is ticked, QMaxent shuffles the values of
each variable in turn and measures the drop in AUC (scikit-learn
permutation_importance, Repeats shuffles per variable, default 10).
The drops are reported as a percentage of the total, the convention of
maxent.jar. The Evaluate on drop-down sets which data the AUC drop is
measured on:
| Option | What is evaluated | Use it to |
|---|---|---|
| Training data (maxent.jar convention) (default) | The full model on all training presences and background points, as maxent.jar does | Compare variable rankings with maxent.jar studies |
| Held-out CV test folds (all folds pooled) | Each cross-validation fold model on its own test fold; the results of all folds are pooled | See how much predictive performance at new locations depends on each variable |
The held-out option needs Geographic K-Fold, Random K-Fold or Checkerboard cross-validation. With no CV or with Buffered LOO, QMaxent falls back to the training data and says so in the log. (v0.1.x used only the first CV fold, which made rankings unstable.) The training log and the model card record which option was used.
Permutation importance is fast (no refitting), but it can underestimate the importance of variables that are correlated with others. Read it together with the jackknife.
Output paths¶
Two file paths default to a qmaxent_output/ subdirectory of the current
QGIS project:
- Model file (.pkl) — the serialised trained model, reloadable from
the Data tab. QMaxent also writes
<model>_model_card.jsonand<model>_coefficients.csvnext to it. See Saving and reusing models for these files and the security note on Python pickle files. - Results workbook (.xlsx) — multi-sheet supplementary table. See Exporting results for the sheet-by-sheet layout.
Clear either field to skip writing that file. The model object remains in memory for the rest of the session even when not saved to disk.
Advanced options¶
A collapsible Advanced section exposes:
- Fix random seed (default on, value 42): one seed for every random
step of a run, that is CV fold partitions, hold-out splits, permutation
shuffles, survey-site shuffling and, since v0.2.0, the background-point
draw (v0.1.x drew a different background on every run). Background
points are drawn from the first raster in the ① Data list, so keep the
raster order the same to reproduce a run exactly. The seed value is
written to the Overview sheet of
results.xlsxand to the model card. - Down-weight spatially clustered points — applies the Phillips et al. 2009 sample-bias correction by giving duplicates and very-near-duplicate occurrences less weight in the likelihood. Recommended whenever the presence layer was assembled from ad-hoc records (museum specimens, citizen-science sightings, etc.).
- Add presences to background — on by default, matching Java MaxEnt convention. Required for the Maxent likelihood to be well-defined (Phillips & Dudík 2008).
Next¶
When everything is configured, click ▶ Run Maxent at the dock footer. Focus shifts to the ③ Training tab.