microsegments

Where do buses and trams linger? microsegments cuts each line into short stretches (30 m by default) and counts how many times a vehicle position is observed in each one, per hour and per day of the week. It works with GTFS plus any vehicle position feed: GTFS-RT VehiclePositions (lat/lon, as CSV, parquet or .pb) or linear-referenced AVL (last stop + distance, like STIB).

Install

pip install "microsegments[plot]"        # plot extra = matplotlib

Python ≥ 3.11. Core dependencies: polars, numpy, shapely, gtfs-parquet (no pandas).

Quick start (offline, synthetic data)

python examples/make_synth.py                     # simulated tram line S1 -> examples/data/
microsegments run examples/synth.toml -o out/     # result.parquet, analysis.json, report.html, matrix_dir*.png

Linear-referenced feeds (STIB vehicle-distance)

[input]
paths = ["vd55/*.parquet"]           # <YYYYMMDD>.parquet (ts, dir, point, dist) + <YYYYMMDD>.snaps.parquet
kind = "stib"                        # or "linear" for any "stop_id + dist_m" table (map columns below)
events = "punctuality/*.parquet"     # optional stop events: passages from both sources, max taken

[input.linear]
id_normaliser = "leading_digits"     # "05766" -> "5766"
feed_length = "auto"                 # rescale feed metres to GTFS link lengths

[gtfs]
path = "gtfs/"                       # zip, .txt directory or gtfs-parquet directory
# dated = "gtfs/{date}"              # one feed per day instead

[select]
route = "55"
dates = "2025-02-17..2025-05-16"
weekdays = [0, 1, 2, 3, 4]

See examples/stib55.toml (runs on the two-day test fixture).

GTFS-RT positions as CSV (lat/lon)

[input]
paths = ["positions/*.csv"]
kind = "latlon"
ts_unit = "s"                        # "s" | "ms" | "us" | "iso"
[input.columns]                      # canonical name = your column
ts = "timestamp"
route_id = "route_id"
vehicle_id = "vehicle_id"
lat = "latitude"
lon = "longitude"
trip_id = "trip_id"                  # optional, helps map matching

Per-vehicle fixes are thinned to one per tick_s (20 s) so counts stay comparable with a snapshot feed. See examples/gtfsrt_csv.toml for every option.

Python

import microsegments as ms

res = ms.run(ms.Config.from_toml("ms.toml"))   # obs -> coverage -> network -> place -> passages
                                                # -> segments -> count -> analyse -> hotspots
res.result            # polars: one row per segment x hour (bands as hour -1 am, -2 pm, -3 day, -4 evening)
res.hotspots          # ranked stretches
ms.export(res, "report.html", lang="fr")       # standalone page (fr / en)
ms.plot.matrix(res, direction_id=0, metric="obs_per_h", days=[0, 1, 2, 3, 4], stops=False)
ms.plot.map(res, 0, "pm", metric="excess_per_passage")
contract = res.contract()                      # the page JSON (also the platform's /api/analysis)

pre = ms.prepare(cfg)                          # reuse placement to tune the segment length
tr = ms.tune.segment_length(pre.placed, pre.segment_fn(), coverage=pre.coverage, passages=pre.passages,
                            pattern_days=pre.network.pattern_days)
ms.plot.tune_curves(tr)

Stages are usable on their own: read_observations, build_network, segment, count, analyse, hotspots, simulate.

CLI

microsegments run CONFIG -o OUT [--lang en] [--segment-m 20] [--dates A..B] [--source "credit"]
microsegments tune CONFIG [--lengths 10,20,30,50] [--sensitivity] [-o OUT]
microsegments hotspots CONFIG [-o hotspots.geojson|.csv|.parquet]
microsegments inspect CONFIG        # coverage per day, GTFS versions, placement / drop counts

What the numbers mean

Every poll (about every 20 s) puts each vehicle in one micro-segment: one observation. One observation is therefore about 20 s of presence.

All values are ratios of sums over the selected days (optionally post-stratified by weekday). Stop zones (30 m before to 60 m after each stop) can be masked to bring out signals and junctions. Colour scales are fixed per metric (p98 over 6–21 h), so hours and weekdays compare directly.

Missing data and GTFS versions

• Hours with < 80 % of polls, a frozen feed (> 5 min), the line absent, or abnormal vehicle counts are dropped from numerator and denominator. Polling gaps are handled by the exposure (covered hours); counts are never imputed.
• Days with more than 25 % of their 6–21 h dropped are excluded and listed with the reason (res.analysis.excluded_days, coverage calendar in the page).
• GTFS changes: patterns are rebuilt per service date; link_key / seg_key are stable across versions. Each segment is aggregated only over the days it belongs to the day's main pattern (the reference too). The page shows the most frequent version; others are selectable. A single feed whose calendar does not cover some dates lends them the patterns of the same weekday.

Choosing the segment length

microsegments tune scores L ∈ {10, 15, 20, 30, 40, 50, 75, 100} m (leave-one-day-out Poisson deviance, split-half reliability, hotspot localisation spread and Jaccard) and picks the smallest L with reliability ≥ 0.8, spread ≤ 30 m and deviance within one standard error of the minimum. 30 m is a good default for 20 s polling at urban speeds; go shorter only with dense GTFS-RT fixes and many days. --sensitivity re-runs phase offsets, gap caps, stop zones, references and passage sources.

Hotspots

Stretches where the excess over the evening has a 95 % bootstrap lower bound above 2 s per vehicle, is positive on ≥ 60 % of days, survives Benjamini–Hochberg (q = 0.1) and lasts ≥ 2 hours or a peak band. Adjacent bins merge, never across a stop / running border. Classes: infrastructure (all day), congestion (peak only), mixed. Needs at least 5 included days. Export as GeoJSON with microsegments hotspots CONFIG -o hotspots.geojson.

License

MIT. The bundled Leaflet CSS is BSD-2-Clause.