From the full decision-support platform
to the micro-application that saves you time

The probability that habitat conditions suit the species, learned from known records and from a fine-grained description of the ground. It is not the probability of finding it on the day you go: a site may be suitable and empty, because the species never reached it, because the habitat has closed up since, or because a determining factor is not mapped. For a species strictly dependent on a host plant, a map that ignores the plant describes that plant's habitat: we say so, and the map is then presented as potential habitat.

Three measures, never just one: ranking power, agreement beyond chance, and an index designed specifically for presence-only data. All three together, because they do not say the same thing. A model can rank very well and still be unusable — the order is right but the probability values mean nothing, and the map must then on no account be reclassified into high / medium / low.

Because everything depends on how validation is done, and two figures obtained differently cannot be compared. Common practice sets aside a randomly drawn share of the records. But two records fifty metres apart resemble one another: the model is then scored on points it already knows, and the score rises without it predicting any better elsewhere. We withhold entire areas and score the model where it has seen nothing. The figure is lower, and it holds. Whoever your provider is, always ask how the validation was carried out.

They are, and it is the first weakness of this kind of model. Records in public databases do not say where the species is: they say where naturalists went — trackside, known sites, designated areas. Without correction, the model partly learns the distribution of observers, and may give major importance to a variable that explains nothing. We draw our comparison points following the real survey effort, reconstructed from all the records available for the area.

A few dozen at the very least; beyond about fifty, the model becomes markedly more robust. Below that, we say so before starting. For a rare species — the most frequent case, since the species at stake in a case are rarely abundant — there are two routes: widen the training set well beyond the study area, or give up on probability and deliver a suitable-habitat map, which does not claim the same thing. The choice made, and the reason for it, appear in the report.

Never, and we will not sell it as such. It targets the survey effort: where to go, at what time of year, what to concentrate the visits on. A species with a single annual generation and a short flight period is easily missed; absence is then concluded, wrongly. The model makes the survey more robust, it does not shorten it.

Two things hold. The validation method, written beside every figure: a performance announced without saying how it was obtained cannot be defended. And traceability: every input is fingerprinted by a digital signature, and the report carries the full register. Three years later, you know which version of which layer produced which figure, and the computation can be repeated identically.

By a blind test, which we offer before any assignment. You choose a site you already know well and you do not send us your records. We build the model, we deliver the map, and only then do we compare. You read the success rate class by class, on your own ground.

What is public, and will stay public: the known limitations of this kind of model, the validation protocols, and the figures together with the method that produced them. That is what allows a piece of work to be judged. The full detail — variables retained, construction, parameters — is disclosed within an assignment or under a confidentiality agreement. And it is disclosed: a result the client cannot have checked by a third party is worth nothing the day it is challenged.

From €300 excl. VAT, depending on the level of proof your case requires — from a rapid exploratory assessment to a full regulatory audit. And on a good many other factors (species concerned, area to be studied, and so on).

Two services are priced individually: invasive species and the thirty-year trend.

In almost every case, the traceability manifest comes with the deliverable: the data used, the thresholds adopted, the limitations of the method, and what the figures do not allow you to conclude.