⑤ Survey Site Selection¶
The fifth tab, ⑤ Survey Site Selection (called Priority Sites for Survey before v0.2.0), turns a habitat-suitability raster into a field-ready list of candidate survey sites. Field surveys usually have one of two purposes: looking for new populations, or checking the model in the field. QMaxent supports both with separate sampling designs grounded in the survey-design literature.
Step by step¶
Step 1. Choose the purpose and the settings¶

- Survey purpose: High-suitability search looks for new populations in suitable, unsurveyed places. Gradient-stratified sampling checks the model along its suitability gradient (Step 2).
- High-suitability search settings: the Minimum suitability and the Sampling order (Random or Top-N (highest first)).
- Sampling strategy: the number of sites and the minimum distances (in metres) from known presences and between sites.
- ▶ Select Survey Sites runs the selection.
Step 2. (Validation surveys) Use gradient-stratified sampling¶

- Select Gradient-stratified sampling.
- Gradient-stratified sampling settings: the Threshold method sets the lowest suitability included. Cells above it are split into four suitability bands and sites are spread over the bands.
Step 3. Addresses and the output file¶

- Reverse geocoding: Add administrative address (OpenStreetMap) writes the address of each site. Address language chooses local names or another language. It needs an internet connection.
- Output: the GeoPackage path and Auto-add to QGIS project.
Step 4. Run and check the sites¶

- The progress bar.
- The status line, for example the reverse geocoding count.

- The status line reports how many sites were selected and where the file was saved.

- The survey sites layer is added to the Layers panel.
Survey purpose¶
Pick one:
| Mode | Goal | Reference |
|---|---|---|
| High-suitability search | Look for new populations in unsurveyed high-suitability areas | Williams et al. 2009 |
| Gradient-stratified sampling | Check the model along its suitability gradient: does observed occurrence follow the predicted suitability? | Rhoden, Peterman & Taylor 2017 |
The two modes ask very different questions of the same map, so they sample it differently:
- High-suitability search (formerly Discovery mode) samples sites only from a band of high suitability, away from known occurrences. These are the most promising places to find the species where it has not been recorded yet.
- Gradient-stratified sampling (formerly Validation mode) splits the cells above a threshold into four suitability quartiles and draws the same number of sites from each. Lower-suitability sites are included on purpose, so the field results can show whether the species is found more often where the model predicts higher suitability.

Mode settings¶
Only the settings group of the selected mode is shown.
High-suitability search settings¶
| Control | What it does |
|---|---|
| Minimum suitability | Cells with suitability at or above this value form the candidate pool. It is filled in automatically as 0.9 × the maximum of the prediction raster when the raster is selected; you can type another value. |
| Sampling order: Random (default) | Draws sites at random from the candidate pool. This gives an unbiased sample within the high-suitability band. |
| Sampling order: Top-N (highest first) | Takes the cells with the highest suitability first. Use it when you want to maximise the expected hit rate. |
Gradient-stratified sampling settings¶
| Control | What it does |
|---|---|
| Threshold method | Sets the lower bound of the sampling frame: 10th percentile training presence (T10; Pearson 2007) (default), Minimum training presence (MTP; Pearson 2007), Maximum sum of sensitivity + specificity (MaxSSS; Liu 2013), or Custom value... |
| Computed threshold value | The threshold computed from the training data for the chosen method (editable only for Custom value...) |
The threshold only defines the sampling frame. The cells above it
are split into four quartiles of suitability, and the same number of
sites is drawn from each quartile
(Rhoden, Peterman & Taylor 2017). No binary
suitable/unsuitable map is produced, and the threshold is not a claim
that cells below it are unsuitable. Each site carries its quartile in
the quartile field (0 = lowest band above the threshold, 3 = highest),
and the layer is styled by quartile. In High-suitability search the field
is always 0.
Spacing constraints¶
The Sampling strategy group is shared by both modes. Besides the Number of survey sites (default 20), two distance fields define the spatial structure of the sampled set:
- Min. distance from existing presences (m) (default 1,000) keeps candidates away from known occurrences. Set it to the species' detection radius (e.g. 1 km for a calling fairy pitta, 200 m for a slug). This prevents the algorithm from re-sampling territory already covered by earlier surveys.
- Min. distance between sites (m) (default 500) keeps the candidate set spatially independent. A typical value is half the distance from existing presences. Spatially balanced sampling, in the sense of Stevens & Olsen 2004, is approximated by setting this value to the nearest-neighbour distance you want in the final set.
QMaxent enforces both constraints exactly: candidates are dropped, not moved, when a constraint is violated, so the surviving set is always within the specification. If the candidate pool cannot supply the requested number of sites under these constraints, QMaxent returns the sites it found and reports the actual count.
Reverse geocoding¶
When Add administrative address (OpenStreetMap) is ticked, QMaxent calls the Nominatim service of OpenStreetMap for each site. No API key is needed. Reverse geocoding works worldwide. It adds four generic OpenStreetMap address levels and the full address:
| Field | Address level | Examples |
|---|---|---|
country |
Country | Dominican Republic, South Korea |
province |
First-level administrative division | a Dominican provincia, a Korean province, a US state |
city_county |
Municipality | a Dominican municipio, a Korean city or county |
district |
Locality | a suburb, village, neighbourhood or hamlet |
display_name |
Full address text returned by Nominatim |
A level that OpenStreetMap does not hold for a location is left blank.
Address language sets the language of the names:
- Local names (default): each country's own names, for example Spanish names in the Dominican Republic and Korean names in Korea.
- QGIS interface language: the language of your QGIS interface, with English as a fallback.
- English: English names where OpenStreetMap has them.
The resulting GeoPackage is directly usable for permit applications and field-team coordination, which matters in jurisdictions where survey access requires prior notice to the local administration.
Requests are limited to one per second to comply with Nominatim's fair-use policy, so 20 sites take about 20 seconds. For 30 or more sites QMaxent first shows the expected time; for large sets, untick the option or select the sites in smaller batches. Results carry the attribution © OpenStreetMap contributors.

Outputs¶
After clicking ▶ Select Survey Sites, QMaxent writes:
survey_sites.gpkg(default name, in the QMaxent output folder): a GeoPackage point layer with the fieldsid,lat,lon,suitability,quartileand the address fields above. With Auto-add to QGIS project ticked, it is added to the map as the layer Survey sites.
If the output GeoPackage already exists, it is replaced, and any layer
of the current project that reads it is removed first. If another
program holds the file open, QMaxent writes to the next free name
instead, such as survey_sites_1.gpkg, and reports the name it used.
The training thresholds (MTP, T10, MaxSSS) that Gradient-stratified
sampling can use are also recorded in the model card
(<model>_model_card.json, see Saving and reusing models).

The sites are immediately ready for export to mobile-GIS apps such as QField, Mergin Maps, or Locus Map for offline field use.
Choosing parameter values¶
Some practical starting points:
- Number of survey sites: 20–30 per species per season is a typical field-survey budget for a single team (Robinson et al. 2018).
- Min. distance from existing presences: the species' typical home-range diameter is a good starting point. Too small and you re-sample known sites; too large and you push candidates into marginal habitat.
- Minimum suitability (High-suitability search): the automatic value (0.9 × raster maximum) gives a focused search; lower it to widen the candidate pool when too few sites satisfy the spacing rules.
- Threshold method (Gradient-stratified sampling): T10 or MTP give a wide frame that includes lower-suitability cells; MaxSSS gives a narrower one.
Worked applications¶
- The Bradypus example walks through both modes at continental scale.
- The Pitta nympha example shows a High-suitability search at municipality scale, with reverse-geocoded administrative addresses ready for field permits.
Next¶
Move to Saving and reusing models to learn how to share the trained model with collaborators or re-project it onto a new raster stack later.