Climbing elevation over-reporting

[lsoules]

BikeChatt is typically reporting about 2x the climb amount compared to Strava. 3 identical ride comparison files attached.

GPXRides.zip

[kwiens]

Doing extensive math on this.

  ┌──────────┬────────┬──────────┬───────────┬───────────────────┬─────────────────┐
  │   Ride   │ Strava │ Original │ After EMA │ After all filters │ Total reduction │
  ├──────────┼────────┼──────────┼───────────┼───────────────────┼─────────────────┤
  │ MTB      │ 36m    │ 126m     │ 87m       │ 117m*             │ -7%             │
  ├──────────┼────────┼──────────┼───────────┼───────────────────┼─────────────────┤
  │ Casual 1 │ 25m    │ 35m      │ 24m       │ 31m               │ -11%            │
  ├──────────┼────────┼──────────┼───────────┼───────────────────┼─────────────────┤
  │ Casual 2 │ 6m     │ 43m      │ 13m       │ 37m               │ -14%            │
  └──────────┴────────┴──────────┴───────────┴───────────────────┴─────────────────┘

The three filters protect against specific failure modes:

• altitudeAccuracy: phones under tree cover or in urban canyons report high uncertainty — now rejected
• Spike rejection: sudden GPS altitude jumps (common during cell tower handoffs) are clamped
• Min distance: altitude oscillations while stopped at a traffic light are no longer counted as elevation gain

We're still significantly inflated — the algorithmic improvements help with specific failure modes (spikes, stopped jitter), but the fundamental issue is that phone GPS
altitude has 5-15m vertical error that no amount of filtering can fully remove. The remaining overreporting is real noise being counted as elevation change.

Additional non-DEM ideas (diminishing returns)

• Barometric altitude via Web Sensor API — AbsoluteOrientationSensor / Barometer aren't widely available on mobile browsers yet. Android Chrome doesn't expose barometer data through the
  Geolocation API.
• Kalman filter — more sophisticated than EMA, but the improvement over EMA+deadband for this noise profile is marginal (~5-10%).
• Aggressive deadband (5-6m) — would help flat rides but would suppress real short climbs on MTB trails.

None of these will close the 3x gap on the MTB ride. DEM is the answer.

[kwiens]

DEM stands for Digital Elevation Model. It's a grid of elevation values covering an area of the earth — basically a topographic height map stored as pixels.

Each pixel represents the ground elevation at that location. Our tiles are z13 (256×256 pixels per tile), where each pixel covers roughly 16×16 meters. The Mapbox Terrain-RGB format encodes elevation into the RGB color values of a PNG image:

height = -10000 + (R × 65536 + G × 256 + B) × 0.1

So a pixel with RGB(1, 131, 144) decodes to:

-10000 + (1 × 65536 + 131 × 256 + 144) × 0.1 = 489.6 meters

This is the same data that makes 3D terrain rendering possible in mapping apps. Strava uses it to replace noisy GPS altitude (which comes from satellites and has 5-15m vertical error) with the known ground elevation at each GPS coordinate. We're now doing the same thing — when a ride ends, we look up each GPS point's lat/lng in our cached DEM tiles and swap in the terrain elevation before computing stats.

The data comes from satellite radar surveys (primarily SRTM — the Shuttle Radar Topography Mission) refined with additional sources. It doesn't change over time since terrain is static,
which is why we can cache it in the repo.

┌─────────────────────────────┬──────────────┬─────────────────┬───────────────┐
│          Approach           │ Ride 1 (MTB) │ Ride 2 (casual) │ Ride 3 (flat) │
├─────────────────────────────┼──────────────┼─────────────────┼───────────────┤
│ Strava (reference)          │ 36m          │ 25m             │ 6m            │
├─────────────────────────────┼──────────────┼─────────────────┼───────────────┤
│ Original algo (5pt, db=3)   │ 126m (+246%) │ 35m (+43%)      │ 43m (+607%)   │
├─────────────────────────────┼──────────────┼─────────────────┼───────────────┤
│ GPS + all filters (current) │ 117m (+221%) │ 31m (+26%)      │ 37m (+519%)   │
├─────────────────────────────┼──────────────┼─────────────────┼───────────────┤
│ DEM + all filters           │ 113m (+211%) │ 31m (+27%)      │ 10m (+57%)    │
└─────────────────────────────┴──────────────┴─────────────────┴───────────────┘

What the data shows:

• Ride 3 (flat): DEM is a huge win — 43m → 10m (vs Strava's 6m). GPS altitude was pure noise on a flat ride; DEM eliminates it.
• Ride 2 (casual): Already close with GPS filters alone. DEM doesn't change it because the ride's GPS positions sample similar DEM pixels. 31m vs 25m is within the DEM pixel quantization noise.
• Ride 1 (MTB): DEM barely helps (117m → 113m). The problem here isn't altitude noise — it's position noise. On an MTB trail under tree cover, the GPS track wanders horizontally, causing the ride to cross different DEM elevation pixels even when the rider is on the same trail. The DEM correctly shows different elevations at those wrong positions, creating phantom elevation changes.

Ride 1 data: it's likely that the app got paused and screwed up the data