Where coqui probably are

The map has a new layer that estimates where coqui are likely to be present. It shows the estimated area in orange, including the locations with coqui reports or recorded calls. If your home is inside the orange area and you hear no coqui, this post is for you.

November 3, 2026 · Tim Storm

The layer is off by default. Turn it on to see the estimated area. Turn it off to see individual reports and listening results.

The map around Haʻikū with the projection layer on: one orange shape over the photo, where coqui probably are, with the lot lines, roads and MISC zones on top and the quiet lots cut out — Estimated coqui area near Haʻikū, shown in orange, from the current run (reports through November 2, 2026). Property boundaries, roads and MISC zones remain visible; properties reported quiet are cut out. Turn the layer off to see reports and listening results.

What the orange area shows

The layer combines coqui observations with information about the land. It estimates where coqui may be present between and near observed locations.

Orange means coqui are probably present. It does not confirm their presence at every location inside the area.

The layer uses one color. It includes both observed and estimated locations. It does not show different confidence levels within the orange area.

Property boundaries, roads and MISC zones (the zones of the Maui Invasive Species Committee) remain visible. Individual reports are hidden while the layer is on. Locations reported as quiet remain unshaded.

I hear no coqui at home. Why is my property orange?

The orange area estimates where coqui may be present. It uses nearby reports, recorded calls, and habitat data such as tree cover, rainfall, and gulches. It can include properties that nobody has reported yet.

In testing, about 9 in 10 checked squares with scores of 0.75 or higher had coqui detections. This result describes the checked squares before quiet locations and open ground are removed from the map. It does not confirm coqui at your home or on a neighboring property.

The model cannot tell when coqui might reach your property. But if your property is orange, coqui have been reported near you and the land suits them. Treat your property as next in line. If it is quiet today, your observation can improve the map.

Mark your property Quiet. Find your address, select Quiet, then Save report. No account is needed. The next published update will leave your property unshaded. This records where you heard no calls; it does not prove that coqui are absent.
Listen again after rain. If you hear coqui, update your report. A newer report replaces the older one.
Respond to the first calls. One or two frogs are cheap to remove. A chorus is not. If you hear coqui, use the How to fight coqui guide to choose your next step.

Your report adds a local observation where the model may have relied on nearby evidence.

How to turn it on

Open the map legend.
Select Layers.
Turn on Projection: where coqui probably are.

The map saves this setting on your device. Select the ? beside the layer to see its update date.

The layer is updated about once a month as new observations become available. The date under the layer name tells you which run you are looking at.

Why this layer helps

Reports and listening results do not cover every location. The layer helps identify places where more observations could improve the map, especially near the edge of the estimated area.

A quiet report inside the orange area leaves that location unshaded in the next published update. This shows where someone has listened without hearing coqui. It helps distinguish quiet observations from places with no reports.

What the layer cannot tell you

It cannot confirm coqui at every shaded location. Some locations have observations. Others are estimates based on nearby evidence and habitat.
It cannot confirm that unshaded areas are free of coqui. An area may be unshaded because of a quiet report, a low model score, limited evidence, or a rule that excludes it.
It does not give a precise result for each property. The model uses squares that are 50 meters by 50 meters. Its grid does not follow property boundaries.
It does not add to the report counts. The legend counts reports and listening results only.
It is not an agency survey or a control plan. Coqui observations come from this site. Environmental data come from the sources listed below.

A quiet report means no calls were reported during that observation. It does not prove that coqui are absent.

The model is updated as new observations become available. Quiet reports also control the display: reported quiet locations are excluded from the orange area until a newer report or detection records coqui there.

How much area is shown?

Map category	Approximate area
Area covered by coqui reports and listening detections	3,200 acres
Total orange area, including those observed locations	6,300 acres

The 6,300-acre figure includes the observed locations. It is not an additional 6,300 acres. The 3,200 acres are the ground that reports and detections cover: a reported small property as a whole, a marked location on a large property, a 100-meter road section with recorded calls, or a gulch section where calls were heard. These figures describe mapped areas, not a count of frogs or a precise measurement of occupied habitat.

The layer went live on November 2, 2026. The current run uses reports and listening results through the evening of November 2, Hawaiʻi time, and the layer shows that date. The ? beside the layer shows the date of the latest run.

Add an observation

Find your address on the map, select Yes or Quiet, then select Save report. No account is needed. These reports train the model: a report from a signed-in neighbor at full weight, a visitor's unverified report at half weight. Turn the projection layer off to see existing reports and listening results.

You can also select I hear a coqui when you hear a call. These taps are shown on the map as “not checked” and are not used to train the projection model.

Open the map with the layer on → I hear a coqui

How the model works

The grid and scores

The model divides the analysis area into squares that are 50 meters by 50 meters. Each square receives a score from 0 to 1. A higher score indicates stronger support for coqui presence.

The orange area uses a score threshold of 0.75, with the exclusions described below. This threshold does not establish a verified 75% chance of coqui presence at each location.

Observations used for training

The model learns from two types of observation:

Coqui detected: a coqui report on a small property, a marked location on a large property, a 100-meter road section with recorded calls, or a gulch section where calls were heard.
Quiet: a property reported as quiet, or a road section checked during calling hours with no calls detected.

Each source has the same total weight, regardless of how many grid squares it covers. An unverified visitor report receives half that weight. I hear a coqui taps are not training observations.

When the model scores a square, it excludes that square's own observation from the evidence used to calculate its score.

Information used for each square

The model uses 18 inputs, including:

Distance to the nearest detection, adjusted for vegetation and terrain. The calculation assumes easier travel through tree cover and along gulch floors, and harder travel across open pasture and lava.
Reports and quiet listening results within 500 meters. Closer observations have more weight.
Tree cover, rainfall, slope, and position within the gulch network.

These inputs describe local conditions. They do not track the movements of individual frogs.

The calculation

The model averages the outputs of two statistical methods: logistic regression and boosted decision trees.

The tree model has a distance constraint. If the other inputs stay the same, a greater distance from a detection cannot increase its estimate.

Each monthly update uses the same 18 inputs and training procedure, with the updated observations.

Testing

Nearby squares can have similar conditions. Testing therefore uses whole blocks that are 1 kilometer across. The model is tested on blocks excluded from training, and a held-out square's own observation is never used to score it.

In block cross-validation, the model had an AUC score of 0.88. This means it ranked a randomly selected detection square above a randomly selected quiet square about 88% of the time. AUC measures how well the model ranks these observations; it is not the percentage of predictions that are correct.

One in five blocks is reserved as a separate test set that was not used to select the model's form or inputs. The AUC score on this set was 0.92 in this run. The same reserved blocks are checked at every monthly update, so this is a standing guard against a bad update, not a one-time final test.

Among observed squares in the held-out blocks, coqui detections were recorded in approximately:

9 in 10 squares with a score of 0.75 or more.
3 in 5 squares with a score of at least 0.4 but less than 0.75.
1 in 4 squares with a score under 0.4.

These rates describe the score threshold alone, before the drawing rules below remove quiet locations and open ground. They apply to observed squares. The observations favor places where coqui are heard. The corresponding rates in unobserved areas are unknown.

Rules for drawing the area

The score must be at least 0.75.
The projection stops at approximately 2.5 kilometers of travel from a detection, measured along the model's cheapest path through tree cover and gulches, not in a straight line. A straight-line distance can be shorter. The cutoff is chosen from the largest spacings between recorded detections (the 99th percentile), so it moves as detections are added.
Open pasture and bare ground are excluded, regardless of score.
Properties, marked locations, road sections and gulch sections reported as quiet remain unshaded, regardless of size. Every quiet report counts, verified or not, and a quiet report does not expire. The exclusion lifts only when a newer report or detection at the same place says coqui.
The published area includes observed locations. Small outlines are smoothed: pieces under 3 hectares and holes under 10 hectares are removed, except holes made by quiet reports. Turn the layer off to see the underlying reports and listening results.

Monthly updates

One command on the site owner's computer retrieves the public data, retrains the model, tests it and creates the layer.

The update is published only if all checks pass:

AUC on the reserved test set is at least 0.88.
At least 85% of observed test squares with a score of 0.75 or more have coqui detections.
The total area changes by less than 25% from the previous published run.
The map shape passes validation.

If a check fails, publication stops for a person to review the result.

Limits and data sources

Where observations are sparse, the model relies more on habitat and distance. For example, estimates for gulches west of Māliko rely on few detections.

The model may miss small areas of suitable habitat, such as a watered garden in otherwise open land. Surveys not included in this site's data may show different results.

Data sources:

Coqui observations: MauiCoqui.com reports and listening results.
Terrain and gulches: USGS 3DEP and the 2020 Maui lidar survey.
Tree cover: Sentinel-2 and NAIP imagery.
Rainfall: Rainfall Atlas of Hawaiʻi.

Lidar uses laser measurements to map land and vegetation. The published layer uses the 50-meter grid and does not name properties or people.