Stream Fish Occurrence Probability Tool (Trial)

How to use it, and what it assumes

• You enter the conditions at a single point in the stream — values measured at each point along a cross-sectional transect.
• It assumes the fish naturally occurs in that river system. In a river where the species is not distributed, the probability has no meaning.
• Water temperature, water quality and connectivity (fragmentation by weirs or dams) are not included.
• To look at the effect of construction, we recommend entering the expected conditions before and after the works and comparing the change in probability.
• If the tool shows “outside the range of the data used to build the model”, the estimate is an extrapolation beyond the data, so treat it with particular caution.

About this attempt

In graduate school I studied where stream fish live in relation to physical conditions such as channel width, depth, current velocity and substrate, and expressed those relationships as statistical models. It was painstaking work: diving to check for fish, catching them to count and measure them, measuring the habitat at the same points, and explaining whether a fish was present or absent from the conditions.

River works, dam construction, road improvements, surveys of protected areas and impact assessments all call for evidence on whether a place is habitable for fish. Yet tools that anyone can use for that judgement are rarely made publicly available.

Perhaps there should be a site that can assess the habitat of living things in this way. Ideally I should continue the research myself and improve its accuracy, but if researchers keep working at it, this kind of assessment should be achievable. This page is a thought experiment — a trial — to test that possibility.

Scale of the surveys

From 2006 to 2009, surveys in rivers in Mie, Gifu and Aichi Prefectures covered about 780 points and plots in total, and about 10,000 fish were recorded or caught (more than 3,700 of them were also measured for body length). Site names and measured values are not published.

For fish, we recorded species identification, presence or absence, number of individuals and body length, and prepared voucher specimens. For habitat, we measured channel width, depth, current velocity, overhanging bank vegetation, emergent plants, submerged plants, overhead cover, stone embeddedness and substrate, and — depending on the survey — water temperature, pH, dissolved oxygen, bank material and elevation. Fish were detected by combining underwater observation with sampling using an electrofisher and various nets.

The tool’s calculations use the data in which fish and habitat were recorded at each point along transects (503 points in six rivers). The other surveys established which fish live in which river systems and which habitat differences relate to differences in fish fauna, and formed the basis for choosing survey sections and the variables to analyse.

Analysis protocol

1. Survey design

Treating a river as a sequence of riffles and pools, we laid cross-sectional transects at regular intervals along each survey section and, at each point on a transect, recorded at the same time whether fish were present and what the habitat was like. Pairing fish and habitat at the same point produces data in which fish occurrence can be explained by habitat conditions.

2. Data integration and quality control

Because recording formats differed between years and rivers, the data were aligned to nine variables common to all rivers, with unified scoring criteria, before being merged. Points with records whose definitions could not be confirmed or with missing values (26 points) were excluded, leaving 477 points from six rivers for analysis.

3. Screening the variables with multivariate analysis

In-stream habitat can be described along axes such as “from riffles with loose stones to pools with accumulated sediment” and “from near the bank to mid-channel”, and the way species partition habitat appears along those axes. The nine variables, which have different units, were therefore first standardised and their interrelationships screened with multivariate analysis. After confirming that no pair of variables overlapped so strongly that one would make the other redundant (all redundancy indicators were within acceptable limits), all nine were kept as candidates for the models.

4. Modelling by variable selection

For each species, we used a statistical model suited to binary presence/absence data and, guided by an information criterion, added and removed variables one at a time to find the combination with the best balance between explanatory power and simplicity. We used variable selection for the following reasons:

• Including all nine variables lets a model learn chance quirks specific to the surveyed rivers (overfitting), which lowers accuracy when it is applied to other rivers.
• Different species respond to different habitat features (riffles with loose stones, bank vegetation, depth and so on), and keeping only the variables each species needs makes the results easier to interpret ecologically.
• It keeps the estimates stable even for fish recorded at relatively few points.

Each species’ model also uses only data from the rivers where that species was recorded, so that absences in rivers outside its natural range are not mistaken for unsuitable habitat.

5. Validation

For fish recorded in more than one river, we built a model from the data without one river — repeating the variable selection from scratch — and predicted the river that had been left out, repeating this for every river (“between-river validation”). Because the variable selection is redone each time, the accuracy is not inflated. For fish recorded in only one river, the data from that river were split into five parts and cross-validated in the same way, repeated five times. Accuracy is reported as the AUC (0.5 = no better than chance, 1 = perfect discrimination).

6. Inclusion criteria and reliability

Fish with an AUC of 0.6 or more are included. In between-river validation, an AUC of 0.75 or more is rated “high”, 0.70–0.75 “medium” and 0.60–0.70 “low”; fish recorded in only one river are labelled “single river”. The tool warns you when an input falls outside the range of the data used to build the model.

Fish included and model summary

Scientific names follow Eschmeyer’s Catalog of Fishes (as of September 2026), except for the Tōkai small spined loach, which is given the subspecies name used in Japan.

Fish with lower reliability

• Kawa-yoshinobori goby: it occurs at almost every point in some rivers and hardly at all in others, so its occurrence rate varies greatly between rivers. Rivers with few absences give the model little to learn about unsuitable conditions, and the relationship with habitat is not consistent between rivers, which likely lowered its discrimination.
• Ayu sweetfish: it swims in schools and moves widely, so the link between the point where it happened to be and that point’s substrate or vegetation tends to be weak. It may also be affected by stocking.
• Pale chub: it is widespread in many rivers and uses both fast and slow water, so habitat differences have a smaller effect.
• Akaza torrent catfish: it is nocturnal and hides under stones during the day, so it is easily missed by daytime underwater observation, and it may actually have been present at points recorded as absent. Having been recorded in only one river also limits the model.
• Gokuraku-haze goby: recorded in only one river and at relatively few points, so accuracy is limited.

For these fish, accuracy could be improved by combining detection methods that miss fewer fish and by surveying more rivers.