Exporting results¶

Every QMaxent training run writes a multi-sheet XLSX workbook alongside the .pkl model. The workbook is intentionally formatted to read as a Supplementary Table block for an academic manuscript: Times New Roman, table-numbered headings, row footers explaining how each metric was computed. You can paste sheets directly into a paper's appendix.

Output folder layout¶

After a training run that also produced a spatial projection and the Save analysis charts as PNG checkbox was on, the output folder looks like this:

QMaxent output folder showing model.pkl, prediction.tif, results.xlsx, priority_sites.gpkg, and the three saved analysis charts as PNG

File	Produced by	Use
`model.pkl`	Training	Reload later via Load existing model (.pkl)…
`prediction.tif`	Spatial projection	Habitat-suitability raster
`prediction.tif.aux`	Spatial projection	QGIS auxiliary metadata
`results.xlsx`	Training	Multi-sheet supplementary table
`priority_sites.gpkg`	Priority Sites tab	Field-ready candidate points
`prediction_roc.png`	Spatial projection	ROC curve (300 dpi)
`prediction_jackknife.png`	Spatial projection	Jackknife bars (300 dpi)
`prediction_response_curves.png`	Spatial projection	All response curves (300 dpi)

Sheet-by-sheet description¶

The XLSX has four to five sheets depending on which features you ran. The screenshots below are from a Pitta nympha run — the format is identical for every dataset; only the values change.

Table 1 — Experimental setup¶

A one-page record of the full run configuration. Reviewers can verify seed, variable count, regularization, CV scheme, and training/CV AUCs at a glance. This is exactly the information Araújo et al. 2019 flag as the minimum methods-section reporting standard for a published SDM.

Table 1 — Experimental setup, training data, and run-level metrics

This sheet is the single best artefact for methods-section reproducibility. Cite it as your "Table S1" and a reader can re-execute the analysis bit-identically.

Table 2 — Predictor variables¶

Lists every environmental variable, its type (continuous / categorical), and its training-data range. For continuous variables the range is [min, max]; for categorical, the discrete set of levels encountered during training.

Table 2 — Predictor variables, types, and training-data range

Use this sheet to document extrapolation scope: any prediction applied to environmental values outside these ranges is by definition extrapolation (Elith, Kearney & Phillips 2010) and should be flagged in your results discussion.

Table 3 — Cross-validation results¶

Per-fold held-out AUC and the across-fold mean ± standard deviation — the headline performance estimate.

Table 3 — Cross-validation results (Random K-Fold)

Folds with degenerate test sets (e.g. all-positive after spatial partitioning) are reported as NaN and excluded from the mean. This matches the way the original Maxent literature handles the rare cases where a fold cannot produce a valid AUC.

Table 4 — Jackknife variable importance¶

For each variable, four AUCs (only-train, only-test, without-train, without-test), plus a "Train AUC drop" and "Test AUC drop" column to make the contribution magnitudes scannable. Variables are sorted by descending Test AUC drop so the most uniquely-informative variables are at the top.

Table 4 — Jackknife variable importance

The footer note explains the drop columns: Drop = full-model AUC − without-variable AUC; larger drops indicate variables whose unique contribution is harder to recover from the others. This is the metric Phillips, Anderson & Schapire 2006 recommend as the most robust indicator of unique contribution.

Optional Table 5 — Priority sites¶

If you ran the Priority Sites for Survey tab, the same workbook gains a fifth sheet with one row per candidate location. The columns include lat, lon, suitability, and — when reverse geocoding was enabled — country, province, city_county, district, and the human-readable display_name. The format is the same one QMaxent uses inside the GeoPackage:

Priority Sites attribute table — 20 candidates with Korean reverse-geocoded administrative addresses

Customizing the export¶

The default file path on the ② Parameters tab is <home>/qmaxent_output/results.xlsx. Change it to anywhere writable; relative paths are resolved against the current QGIS project folder.

To stop the XLSX from being written (rare — usually faster to ignore the file than to disable it), simply clear the path field. QMaxent will skip the export and emit a notice in the training log.

The PNG analysis charts are controlled by the Save analysis charts as PNG checkbox on the ④ Results → Spatial Projection sub-tab. With it ticked, three additional PNGs are written next to the GeoTIFF at the moment you run projection. They are sized for direct paste into a single-column manuscript figure (300 dpi, ~1.6 MB each).

Citing the workbook¶

The XLSX was designed to be cited the way published supplementary tables are cited:

Yu, B.-H. (2026). QMaxent results workbook for [Supplementary table]. https://github.com/osgeokr/qmaxent

If your model is reproducible (fixed seed + same input rasters) you can also distribute the model.pkl alongside the XLSX so reviewers can re-load the model and inspect predictions independently.