Files
aimsight-labeling/auto/analyses/WYOMING_2026-11-06T0805_2026-11-16T0627/analysis.md
T

79 lines
17 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Labels: WYOMING, 2026-11-06T08:05Z → 2026-11-16T06:27Z (UTC)
Measurements of the range: 2026-11-06T08:15Z → 2026-11-16T06:15Z, 15 min, 951 instants. 16 sensors: 3 references (RD, target_12, target_13), 13 targets. No image quality metrics (export columns peak to snr empty). One image per hour, on the hour. Temperature: only `temperature_deg_c` is filled; I call it the housing temperature (`labeling-rules.md` §5). No earlier labels (`labeling/auto/WYOMING.jsonl` does not exist).
**Revision after the user's remark on the weather.** I added the MeteoSwiss hourly weather at Genève-Cointrin (station GVE, 4.9 km from "Genève", the place given for WYOMING in `~/aimsight-data/extraction/stations.txt`). I fetched it with `python3 weather/meteoswiss.py "Genève" 2026-11-05 2026-11-16T07:00` from `~/aimsight-data`, into `work/weather_geneve.csv` (05/11 → 16/11 07:00, for the scripts). Since the instructions now require it, the weather of the exact range is also in `work/weather.csv` (`python3 weather/meteoswiss.py "Genève" 2026-11-06T08:05 2026-11-16T06:27`; stderr: station GVE Genève / Cointrin, 4.9 km, 411 m; 238 hours). It is the weather a few km away; the images prevail. What it changed:
1. The housing temperature follows the **outside air** (r = +0.82, no lag) and not the sun (r = +0.02 with radiation, −0.11 with sunshine). Under the bise (13–14/11) the housing cools more than the air, and the shift of the nearest prism follows the housing then too (r = +0.96 with the housing, +0.83 with the air, best at lag 0 h). The moving element therefore follows the camera's own temperature within the hour, which a massive pier would not. This strengthens "thermal movement of the camera mount" for the 8 cycles.
2. (Withdrawn after the expert's review: the agitation episodes are no longer labelled, see §1.) The agitation of the far prisms follows the **cold of the air** (r = −0.63 with the air temperature, +0.05 with the wind). It hits only the long, low sight lines over the river. I now label it apart, as 2 **Visibility** episodes (turbulence over the warmer water, "consistent with"). The one-measurement jump of target_12/13 at 14/11 21:00 is the strongest moment of the bise episode and is named in it; the one of 07/11 05:00 has its own Displacement label again since the 07/11 morning episode was removed.
3. The two dawn transients (08/11, 14/11) fall at the coldest air of a clear night, at sunrise, inside the agitation. A bending of the long sight lines over the river now competes with a fast camera movement. Neither is established, so I moved them from Thermal displacement to **Displacement** (cause not established, both hypotheses stated).
4. The occultation of target_14 (11/11) covers the rain of that morning and ends soon after it stops: water on the prism is likely, and the description says so. The 11/11 07:15 jump falls at the start of the rain, which is now cited as possible. For the 11/11 14:45 passage the rising gusts were checked: the nearer prisms do not move in the raw image, which rules out a camera shake.
**The key to the hint (unchanged).** RD has no spot of its own. Its position falls on a dark part of the nearest pier, next to target_01's prism (`fig_rd_target01.png`). RD and target_01 have the same `error_value` at every instant and the same raw movement (`overview.py --rigid`), so RD is a pattern on target_01's prism: twins, the nearest prism. target_12 and target_13 are the farthest prisms. With temperature the nearest prism moves sideways against all the others, and the similarity fitted on RD + target_12/13 spreads this over the image. target_14, target_15 and target_02, which lie in the image between RD and the far references, take it in their corr_x. target_01, being RD's own prism, does not.
## 1. My labels
### Thermal cycles: 06/11 09:15 → 07/11 07:30; 07/11 07:30 → 08/11 07:00; 08/11 07:00 → 09/11 09:45; 09/11 09:45 → 10/11 04:45; 10/11 04:45 → 13/11 09:30; 13/11 09:30 → 14/11 07:00; 14/11 07:00 → 15/11 04:15; 15/11 04:15 → 16/11 06:15
- **Phenomenon of the catalogue**: "The camera drifts away when the temperature of its housing drops, without a daily cycle" / "The camera moves through a thermal effect". The poor compensation is a new variant: the reference network mixes the nearest and the farthest prisms, so a small movement of the mount becomes a parallax that passes into the targets between them.
- **Precedents**: CLARIDEN-5, -13, -27, -43 (reviewed): a sheltered station without a daily cycle, cut at the turning points of the housing temperature. NORDEND-83/88 (reviewed): "thermal cycle of the camera, poorly compensated", outward and return in one label, labels following each other without rest.
- **Family**: Thermal displacement. The shift follows the housing temperature and comes back with it at equal temperature: dx_RD +0.47…+0.58 at 26.8–27 °C on 07/11, 09/11, 10/11 and 11/11, and −0.05…+0.05 near 24.4 °C on 06/11, 08/11 and 13/11, with up to ~0.3 px of hysteresis. It follows the housing better than the air, at lag 0, and the housing follows the outside air, not the sun.
- **Measured**: `metrics.py` 05/11 → 16/11: r(temperature, cam_scale) = +0.97, r(cam_pan) = −0.94, r(tgt_corr_y) = +0.84. `overview.py --rigid 2026-11-06T09:00Z`: RD and target_01 dx −0.86…+0.73 px. target_02, target_14 and target_15 move less than ±0.41, and the far targets less than ±0.3. `scripts/table.py`: corr_x of target_14 from +0.69 (14/11 06:00) to −0.66 (11/11 12:00), target_01 within ±0.16. Turning points at 15 min: 06/11 09:15, 07/11 07:30, 08/11 07:00, 09/11 09:45, 10/11 04:45, 13/11 09:30, 14/11 07:00, 15/11 04:15. 10/11 → 13/11 stays one label: a plateau on 11/11 and a flat stop on 12/11, with no change of direction. `scripts/weather.py`: r(dx_RD, housing) = +0.96, r(dx_RD, air) = +0.83, r(housing, air) = +0.82, r(housing, radiation) = +0.02. The lag of dx_RD on the housing is best at 0 h. Weather of each cycle (MeteoSwiss): grey and mild 06/11; sunny midday then a clear calm night near 0 °C 07–08/11; sunny morning 08/11; mild, overcast, rain at times 10–12/11; bise (NE, gusts to ~50 km/h, dry air) from the night of 12–13/11 to 14/11; SW wind and mild air 15/11.
- **Images**: 06/11 12:00 whole image (`fig_scene_20261106T1200Z.jpg`); zone thumbnails 06/11 12:00, 14/11 03:00 (`fig_rd_target01.png`); far piers 07/11 07:00, 08/11 06:00/07:00/08:00, 11/11 07:00/08:00/14:00/15:00: nothing in front.
- **Confidence**: high on the family and on what moves; medium on the exact cut points (±1 measurement) and on the mechanism (camera mount most likely).
### Visibility: none (all removed, according to the expert)
- I had labelled four episodes of agitation of the far prisms in cold air (06/11 09:45 → 15:00; 07/11 02:45 → 10:00; 07/11 20:45 → 08/11 13:00; 13/11 02:45 → 15/11 14:00). The expert sees nothing in the data or the images; all four are removed. The statistics behind them stay in `scripts/agitation.py` (hourly median of |corr_y step| of target_07…11 a few hundredths of a pixel higher in cold air, r = −0.63 with the air temperature, nothing with the wind). They are a correlation, not a disturbance seen on the curves.
- The two one-measurement jumps of target_12/13 that I had merged into these episodes have their own Displacement labels again: 07/11 04:45 → 05:15 and 14/11 20:45 → 21:15.
### Displacement: 08/11 06:45 → 08:30 and 14/11 07:00 → 08:30 (dawn transients)
- **Phenomenon of the catalogue**: none fits exactly. Close to "A step of the camera that the compensation does not erase", but this one returns.
- **Precedents**: none with this signature.
- **Family**: Displacement (cause not established). Both come at dawn after the two coldest nights, at the coldest air hour, and end as the sun comes out (sunshine from the 08:00 hour at Cointrin both days). Two hypotheses remain: (a) a lateral bending of the long sight lines through the cold air over the water, since all the far prisms shift together and the near one does not; (b) a fast camera movement as the housing warms abruptly (+1.4 °C and +0.8 °C in the 07:00 hour, with no change in power or load in `metrics.py`).
- **Measured**: `scripts/far.py`: far prisms 03–13 raw x −0.51/−0.67/−0.38 at 08/11 07:15/07:30/07:45, back at 08:00. corr_x target_14 +0.36 (06:45) → −0.75 (07:15) → −0.09 (08:15) → +0.14 (08:30). On 14/11: far x +0.32 → −0.30 (07:45) → +0.06 (08:00); corr_x target_14 0.68 → −0.04 (07:45) → 0.40 (08:30).
- **Images**: 08/11 07:00 (haze over the river, sun on the facades), 14/11 07:00 (sun on the facades). No image during the shift.
- **Confidence**: high on the event, low on the cause.
### Displacement: 11/11 07:00 → 07:30
- target_12/13 jump alone, `error_value` 0.26/0.21 against ~0.08; corr_y of 03–11 −0.08…−0.18; back at the next measurement. The air is mild (11.7 °C, not a cold-air agitation hour); the rain begins in that hour. Cause not established (rain veiling the far references, or a partial passage). Images 07:00 and 08:00: nothing in front. Confidence: high on the event, low on the cause.
### Obstruction: 07/11 13:00 → 13:30, 15/11 14:00 → 14:30, 11/11 14:30 → 15:00
- "A pass-by hides the targets for the length of one measurement", deduced from the signature: one measurement, a group rejected, points jumping, all back at the next. On 11/11 14:45 everything beyond a given pier is lost (target_07–13, references included: no pose, the corr of 02–06 meaningless). The raw positions of the nearer prisms barely move (target_06 ≤ 0.45 px), which rules out a camera shake by the gusts rising then (Cointrin 17.6 → 28 km/h). A crossing of the line of sight, such as a boat on the river, fits that geometry but is not seen. On 07/11 and 15/11 there was some wind (gusts 18 and 33 km/h); nothing else moved. Confidence: medium (no image at the instants).
### No data: 10/11 08:15 → 09:00
- 08:30 and 08:45 are missing for all sensors, and ALASKA and NADELHORN miss exactly the same instants. WYOMING `last_sync` lists 08:45, BusVoltage is 14.1 V, and the images at 08:00 and 09:00 exist. Common cause outside the station, not established. Confidence: high.
### Obstruction: 11/11 09:30 → 14:15 (target_14)
- "An obstacle covers part of the spot". On the 12:00 thumbnail part of target_14's spot is darkened; it is whole at 07:00, 10:00, 13:00 and 15:00, and target_15 is always intact (`fig_target14_occultation.png`). `error_value` 0.077 → 0.252 (12:00) → 0.071 (13:45). corr_x(14) − corr_x(15) −0.32 at 12:00; corr_y −0.21 at 12:00, back by 14:15. Rain at Cointrin 07:00–12:00 (1.9 mm at its heaviest, in the 09:00 hour) and none from 13:00, matching the rise and the end: water on the prism is likely. Family Obstruction, since a single spot is hidden (`families.md`: "under a single spot, it is an obstacle"). Confidence: high on the occultation, medium on the water.
## 2. What I left without a label
- 06/11 09:45–15:00: a slight agitation of the far targets (hourly |corr_y step| 0.025–0.040 at 10:00, 11:00, 14:00, against 0.010–0.020 around it), air 8–10 °C. Labelled Visibility in my second version, removed: according to the expert, nothing shows in the data or the images. Too weak to be told from the ordinary noise.
- 14/11 10:15: target_12/13 opposite raw jumps and corr_y of 07–10 +0.10…+0.14 for one measurement; too small to stand out from the noise of these days.
- The agitation of the far targets in cold air (07–08/11, 13–15/11): a few hundredths of a pixel more from one measurement to the next; not labelled, according to the expert.
- 09/11 night (air 8.1–8.8 °C): agitation of the far targets 0.02–0.03, just above the mild-day level. Below what I bounded as episodes; borderline.
- 10/11 00:00–09:00: the raw y steps of target_12/13 rise during the night rain (up to 0.18 at 03:00) while the far targets' corr stay calm; nothing on the curves.
- `error_value` flags on target_14/15 alone (07/11, 08/11, 15/11), with corr within ±0.07; the slight rise of target_14's `error_value` from 10/11 10:45 to 11/11 08:00, without corr effect.
- The camera pose is missing from 16/11 00:30 to the end while the corr go on: an export matter.
- Tiny one-step deviations of RD/target_12/13 corr (references, own noise).
## 3. The documentation
- **Indications from the expert**: 06/11 09:45 → 15:00, no disturbance in the data or the images; 07/11 02:45 → 10:00, too sensitive; then all the Visibility agitation labels removed, the expert seeing nothing in the data. Lesson: an agitation of a few hundredths of a pixel, found by statistics (steps per hour, correlation with the air) and not visible on the curves, is not labelled. The expert also questioned the thermal cycle 06/11 09:15 → 07/11 07:30. I kept it: corr_x of target_14/15 move by 11–15 noises, the same signature as the other cycles in smaller, and `labeling-rules.md` §1 (D-117, 5-noise rule) asks for it; it awaits the expert's decision. Lesson: an agitation episode is labelled only when it stands clearly above the noise on the curves themselves, not on a correlation with the air temperature alone.
- **Indication from the user**: the weather data. It was not in the station folder, nor in `tools/GUIDE.md`, `tools/DATA.md` or the labeling knowledge, so I had not looked for it. It sits in `~/aimsight-data/weather/meteoswiss.py`, with the place of each station in `~/aimsight-data/extraction/stations.txt`, and is shown by the Data page. `tools/GUIDE.md` should get a weather tool or a line (air temperature, rain, wind, sunshine at the nearest MeteoSwiss station). `reading-rules.md` §3 says the housing temperature is "a witness for the lens and the mount, not for the structure nor the air". Here it follows the air closely (r = +0.82) except under the bise, which is worth recording.
- **What helped**: `families.md` "Thermal displacement or Displacement" (return at equal temperature); the CLARIDEN tunnel precedents for cutting without a daily rhythm; `reading-rules.md` §2 and §5 (`--rigid` first); `tools/DATA.md` (`x, y` are not the centroid), which made "RD is a pattern beside its prism" plausible; `reading-rules.md` §6 "a twin of a reference".
- **What was missing**: a WYOMING row in `stations.md`. It should say: a bridge seen from beneath its deck along a river, in Genève; targets on the piers; RD the twin of target_01 (same prism, nearest pier); target_12/target_13 the farthest prisms; the targets between them in the image (target_02, target_14, target_15) take the thermal parallax; only `temperature_deg_c`; one image per hour. Two phenomena are missing from the catalogue: "references at very different depths, so a small movement of the mount becomes a deformation passed into the targets between them", and "agitation of the long sight lines over water when the air is colder than the water, without the `error_value` rising".
- **Contradicted or unclear**: the catalogue line "Atmospheric turbulence… `error_value` up", and the boundary "without the `error_value` rising, an agitation… is the camera heating". Here the agitation is in the air (it follows the cold air, only on the long sight lines over the water) with a flat `error_value`. I chose Visibility against the letter of that boundary and said so in each label.
- The hint "parallax does not fit target_01": target_01 is RD's own prism, so it moves with the reference and cannot show the parallax that RD passes into the others.
- **Instructions changed during the session** (the weather fetched every time; labels detailed for retrieval; a short report in plain words on an imposed outline, with a GIF when the cause is seen over time). Labels kept as they were (already detailed); report rewritten on the new outline, with a GIF of target_14 on 11/11 07:00 → 15:00 (`target14_occultation_20261111.gif`, zone 1990,725,300,80) and the main figure redrawn in plain words.
## Difficulties
- `tools/GUIDE.md` §Conventions limits Bash to the tools without scripts or loops, while the request allows scripts; I followed the request.
- The weather source was not documented in my instructions at the start; the user pointed to it, and the instructions now require it (section "The weather").
- Hourly images only, no image quality metrics: the brief events have no image at their instant.
## Sources read
`instructions/analyst.md`; `tools/GUIDE.md`; `tools/DATA.md`; `labeling/families.md`; `labeling/labeling-rules.md`; `labeling/reading-rules.md`; `labeling/stations.md`; precedents (`labeling/scripts/precedents.py`, three queries); `request.json`; WYOMING data through `inventory.py`, `overview.py` (default and `--rigid`), `metrics.py` (temperatures; power, current, load), `summary.py`, `ranges.py --rolling 4 --instants 2`, `thumbnails.py`, whole images 06/11 12:00, 08/11 07:00, 14/11 07:00; raw CSVs of ALASKA, NADELHORN and the other stations on 10/11 08:00–08:49; WYOMING `last_sync`, `BusVoltage`; the MeteoSwiss weather (`~/aimsight-data/weather/README.md`, `meteoswiss.py`, `extraction/stations.txt`, `app/server/station.py` for the link); my scripts `scripts/table.py`, `far.py`, `instants.py`, `weather.py`, `agitation.py`, `figure.py`, `figure_weather.py`.