Methodology — level crossings

1. The question we model

Each level crossing on a rail network sits inside a road graph. Around it sit emitters — establishments whose activity generates risk-bearing traffic — and escape targets — graph nodes toward which that traffic plausibly flows. Public data answers three questions about this scene cleanly: where each level crossing sits, where each emitter sits, and what the road graph around them looks like.

One question public data does not answer:

• Which emitters' nominal routes actually cross which level crossing — and therefore which emitters depend functionally on which level crossing, and how acutely.

That missing piece is what SAMRoute imputes. The imputation runs crossing by crossing, across the whole national portfolio, against the same reference, replacing the qualitative presence–absence annotation that current sector practice relies on.

2. The h0/h1 framing

For every (emitter → escape target) pair around a level crossing, two routes can be computed on the road graph:

• h0 — the nominal route on the road graph between the emitter and the target
• h1 — the best alternative route between the same emitter and target that detours around the level crossing

The imputation rule is simple: if h0 crosses the level crossing, that emitter's flow depends on it. The pair (emitter, target) is then recorded as an origin–destination (OD) couple attached to the crossing. The comparison between h0 and h1 quantifies how acutely the flow depends on the crossing:

• SPOF — single point of failure: the OD couple has no h1 alternative; the level crossing is structurally required for the flow to exist on the road graph
• Prohibitive detour — h1 exists, but the cost ratio h1/h0 exceeds an operationally meaningful threshold; the level crossing is functionally required even though a graph alternative exists

If h0 does not cross the level crossing, the (emitter, target) pair is independent of it in routing terms and contributes nothing to its dependence profile.

On the current French national portfolio, roughly 40% of level crossings carry at least one OD couple in the prohibitive-detour class.

3. The pipeline in seven steps

The same pipeline runs for every crossing on the network:

1. Identify at-risk emitters in the surrounding zone — SEVESO sites, logistics hubs, schools, elderly care homes, hazardous-trade and hazardous-waste establishments.
2. Enumerate plausible routes from each emitter toward escape targets in the surrounding road graph.
3. Detect routes that cross the level crossing — these are the h0 candidates.
4. Search for alternatives that detour around the crossing — these are the h1 candidates.
5. Measure the marginal effort of the detour: extra time, extra distance, ratio h1/h0.
6. Evaluate the relevance of h1 per OD couple — SPOF (no h1) or prohibitive detour (h1/h0 above threshold).
7. Qualify the level crossing on the basis of the OD set: how many SPOFs, how many prohibitive detours, against which emitter mix.

The output of step 7 is what the application surfaces, both at portfolio level (rank and filter the whole network) and at per-crossing level (every OD pair, with its h0, h1, ratio, and indicator flags).

4. The metric families

Every level crossing is described by a vector of roughly 200 metrics, grouped into five families that together characterise its nearby environment:

• Population and built environment — population density, average building height, settlement footprint
• At-risk emitters — establishments selected from the SIRENE national business registry against a curated set of activity codes from the French standard industrial classification (NAF), covering transport (rail and road freight, including dangerous-goods carriers under ADR for road, IMDG for sea, and ADN for inland waterway), logistics and warehousing, chemicals and fuels, hazardous-goods trade, and hazardous-waste handling
• Sensitive sites — schools, elderly care homes, SEVESO sites, and other establishments listed in the EPSF audit guide for level crossings
• Escape targets and road-network topology — graph nodes that absorb redirected flow, together with the local road-graph properties that govern h0/h1 routing
• Routing economics — h0 and h1 metrics (time, distance), the marginal detour ratio, the SPOF and prohibitive-detour indicators

The selection of NAF codes is built by substitution: a standardised "at-risk emitter" registry does not yet exist, so NAF activity codes serve as a surrogate indicator, with the curated subset published as part of the model documentation.

5. Coverage, freshness, and known limits

The current production scope covers 9,300 of the approximately 12,000 active level crossings on the French national rail network. The 2,700 missing crossings are excluded by an inclusion criterion that is identified and addressable; full-network coverage is the next planned iteration, agreed with SNCF Réseau's level-crossing safety team.

The data underpinning the model is a snapshot frozen at the time of analysis. Continuous refresh against live data feeds is the next operational step on the production roadmap. Open-data availability outside France is uneven (for example, the Spanish business registry is not geolocated as of today), which paces the geographic extension to other networks.

6. Reproducibility

Every per-asset score unfolds back to the layers, assumptions, and reference materials that produced it. The same inputs produce the same outputs. Source layers, version stamps, and the curated NAF subset are documented per release; the dataset catalogue is exposed at api.samroute.com/api/public/data-catalog.

7. Reach us

For methodology questions, technical due diligence, or model-card review, write to contact@samroute.com. We route methodology threads to the right person directly.