WebCam254

WebCam254 is a time-lapse image dataset of 548,376 JPEG frames from 254 public outdoor webcams, recorded simultaneously every 120 s for 72 h (2026-08-25 23:23 UTC to 2026-08-28 23:21 UTC). 247 of the cameras belong to the ALERTCalifornia wildfire-detection network; seven are weather and community webcams in Nevada, New Mexico, Missouri and Jalisco (Mexico).

Every frame comes with a systematic English caption and a structured record: the exact position of the Sun and the Moon at the image time (including the partial lunar eclipse of 28 August 2026), local time and light phase, the verified camera view (many cameras pan, tilt and zoom), pixel statistics, detection of stale (repeated) images, and validated machine estimates of cloud cover and sun glare. Every statement in a caption can be traced to one of these fields.

At a glance

Item	Value
Frames	548,376 JPEG images (181.2 GB)
Cameras	254: ALERTCalifornia 247, Albuquerque Soaring Club 2, NVWX Elko weather 2, O'Fallon weather station 2, Ajijic Weather 1
Locations	California 247, Nevada 2, New Mexico 2, Missouri 2, Jalisco (Mexico) 1
Capture	every 120 s for 72 h = 2,160 scheduled requests per camera (548,640 in total); 249 failed (network errors), 15 skipped because the previous request overran
Frames per camera	2,154 frames: 1 camera; 2,157 frames: 3 cameras; 2,158 frames: 24 cameras; 2,159 frames: 201 cameras; 2,160 frames: 25 cameras
Image sizes (pixels)	1920 x 1080 (537,591 frames); 3840 x 2160 (4,317 frames); 2560 x 1440 (2,157 frames); 1280 x 720 (2,157 frames); 2560 x 1920 (2,154 frames)
Camera views	364 verified views; 76 cameras used more than one pan/tilt/zoom framing
Frames inspected individually	3,499 (each with its own visual caption)
Hub format	184 WebDataset shards of about 1 GB (183.7 GB) + metadata/frames.parquet
Language	English
Licence	annotations CC BY-NC 4.0; images remain subject to the camera operators' terms (see Licensing)

Quick start

from datasets import load_dataset

# stream without downloading everything (the full dataset is 184 GB)
ds = load_dataset("ezharjan/WebCam254", split="train", streaming=True)
sample = next(iter(ds))
sample["jpg"]                                      # PIL image
sample["txt"]                                      # caption
sample["json"]["illumination"]["sun_altitude_deg"]  # full record (see "Record fields")

Metadata only (all 548,376 records, no images, 80 MB), e.g. to select frames before downloading shards:

import pandas as pd
from huggingface_hub import hf_hub_download

meta = pd.read_parquet(hf_hub_download("ezharjan/WebCam254", "metadata/frames.parquet", repo_type="dataset"))
night = meta[meta.illumination.map(lambda x: x["light_phase"]) == "night"]
shards = sorted(meta.loc[meta.camera_id == "atlas_peak", "webdataset_shard"].unique())   # shards holding one camera

With the webdataset library:

import webdataset as wds
url = "https://huggingface.co/datasets/ezharjan/WebCam254/resolve/main/data/train-{00000..00183}-of-00184.tar"
ds = wds.WebDataset(url, shardshuffle=False).decode("pil").to_tuple("jpg", "txt", "json")

Dataset structure

Repository layout

README.md                  this dataset card
LICENSE                    licence terms (CC BY-NC 4.0 for annotations; image rights, see below)
data/train-XXXXX-of-00184.tar   184 WebDataset shards, ~1 GB each, ordered by camera and time
metadata/frames.parquet    one row per frame: the full record plus `webdataset_shard` and `webdataset_key`
cameras.json               per-camera metadata: location, time zone, capture statistics, verified scene and views
captions.json              all captions grouped by camera (compact JSON)
captions_schema.json       JSON Schema of the frame record, captions.json and cameras.json
scripts/hf_publish.py      the script that built, verified and uploaded this repository

Samples

Each frame is one WebDataset sample with three members that share the key <camera>__<YYYYMMDD_HHMMSS_NNNNN> (the camera folder name lower-cased with every non-alphanumeric run replaced by _, e.g. mt_vaca_8_south__20260826_120331_00381):

• <key>.jpg - the original JPEG, byte-identical to the downloaded file (its MD5 is json.image.md5);
• <key>.txt - the caption;
• <key>.json - the complete frame record described below.

YYYYMMDD_HHMMSS is the scheduled request time in UTC and NNNNN the camera's running request number (1 to 2,160; failed requests also use a number, the 15 skipped slots do not).

Record fields

Field	Meaning
file_name, path	image path relative to the camera folder / to the dataset root of the unpacked edition, e.g. atlas_peak/frames/20260826_024331_00101.jpg
camera_id, camera_name, frame_index	camera folder name, display name, request index from the file name
caption	the systematic caption (see Captions)
time.request_utc, download_completed_utc	scheduled request and download-complete time from the capture log
time.overlay_exposure_utc	exposure time read from the burned-in overlay clock when available (ALERTCalifornia), else null
time.image_time_utc, image_time_source	time used for all astronomy: overlay_clock_ocr (overlay read with confidence >= 0.9) or capture_request_time
time.local_time, timezone, utc_offset	local wall-clock time at the camera (ISO 8601 without offset), IANA zone, offset
illumination.light_phase, day_period	day, low_sun, civil_twilight, nautical_twilight, astronomical_twilight, night; morning/evening (null for day and night)
illumination.sun_*, moon_*	topocentric apparent altitude and azimuth in degrees (azimuth from north through east), lunar illuminated fraction and phase name
illumination.lunar_eclipse_phase	penumbral or partial_umbral while the Moon is above the horizon during the eclipse, else null
camera_pose_overlay	pan/tilt/zoom and OCR confidence read from the image overlay (null for cameras without such an overlay)
view.view_id, assignment, short_description, template_correlation	verified view shown (v1, v2, ... per camera; null if unassigned), how it was established, its short description, thumbnail correlation with the view's daytime template
image.*	pixel measurements: width, height, md5, mode (colour, monochrome, black), exposure (normal, very_dark, overexposed), luminance mean/std/percentiles, dark and clipped fractions, mean RGB, chroma, colourfulness, Laplacian sharpness, luminance by thirds, 8x8 difference hash, mean absolute change from the previous frame
file_bytes, file_bytes_equal_to_previous	JPEG size; whether it equals the previous frame's size
repeated_image_of	earlier frame with byte-identical content (stale feed), else null
flags	repeated_image, identical_to_previous_frame, low_similarity_to_view_template, overlay_clock_repeated, overlay_clock_offset_gt_15min, jpeg_truncated
condition_estimates	machine estimates for colour frames with the sun above the horizon that are not repeated images, else null: cloud_cover (clear_or_few_clouds, partly_cloudy, mostly_cloudy_or_overcast, or null when not confident) with cloud_cover_probabilities; sun_glare_or_flare (true, false or null when not confident) with its probability
visual_check	for individually inspected frames: a visual caption, sky condition, visible sun/moon, smoke plume, image mode, issues, source (contact_sheet_tile, ptz_state_check, unassigned_cluster_check, condition_label_sample) and, for condition-label samples, the full labels; else null

captions_schema.json gives the typed definitions; cameras.json adds per camera: coordinates, time zone, network and partner tag, image URL, capture statistics (scheduled slots, failures with time and error, skips), the site summary, overlay and privacy-mask notes, night imaging mode, every verified view (description, short_description, horizon type, approximate sky fraction, land cover, notable objects, camera hardware in view, frame count, merged_view_ids, contact-sheet tiles), the nine contact-sheet frames (reviewed_tiles; "tile N" in reviewer_notes and overlay notes refers to them) and the light-phase counts.

Splits

A single train split holds all frames. No official evaluation split is defined. Frames of one camera are strongly correlated, so held-out evaluation should split by camera_id.

Unpacked edition

The dataset was built in an unpacked layout (<camera_id>/frames/*.jpg, <camera_id>/metadata.jsonl with one record per line, <camera_id>/manifest.jsonl and session.json capture logs, and the root files above). scripts/hf_publish.py pack converts it deterministically into this repository and verify re-checks every image against its MD5.

Captions

Captions are generated deterministically from the record, so equal facts are always worded identically:

<Light phase> frame from the <camera> camera[ (<network>)]: <verified view description>.
Sun <just above the horizon | low | moderately high | high> in the <direction> (<a>° altitude) | Sun <d>° below the horizon
    [; <phase> moon (<p>% illuminated) <a>° up in the <direction>[, in partial lunar eclipse (umbral phase)
                                                                | , in the penumbral phase of the lunar eclipse]]
    [; moon below the horizon].
<Colour image | Monochrome image (infrared/night mode) | Image almost entirely black>[, very dark | , largely overexposed].
[Estimated conditions (automatic): <clear sky or few clouds | partly cloudy sky | mostly cloudy or overcast sky>[; sun glare or lens flare].]

• Light phase: Daytime, Morning/Evening low-sun, civil-twilight, nautical-twilight, astronomical-twilight, Night.
• Sun wording: just above the horizon < 2°, low < 20°, moderately high < 45°, high >= 45°; directions use eight compass points. They describe where the Sun or Moon is in the sky, not where the camera points.
• The Moon is mentioned when the Sun is below 6° (and "moon below the horizon" when the Sun is below -6°); moon phase names follow the Sun-Moon elongation (full: 168.75°-191.25°).
• Frames whose view could not be established say "camera view not visually verified for this frame" (or give the burned-in PTZ position followed by "view not visually verified"); views whose framing could not be determined from their sample say so after a semicolon.
• Repeated (stale) images say so, name the earlier frame they duplicate and state that the image does not show the conditions at request time; they receive no condition estimates.
• Machine estimates are always introduced by "Estimated conditions (automatic)".

Examples:

• big_rock_ridge/frames/20260827_020331_00801.jpg

  Daytime frame from the Big Rock Ridge camera (ALERTCalifornia): grassy oak hills, dirt roads and a developed valley below a large mountain. Sun low in the west (7.9° altitude). Colour image. Estimated conditions (automatic): clear sky or few clouds; sun glare or lens flare.

• allen_peak/frames/20260827_162131_01230.jpg

  Daytime frame from the Allen Peak camera (ALERTCalifornia): dense conifer forest and layered ridges with grassy patches. Sun moderately high in the east (31.5° altitude). Colour image. Estimated conditions (automatic): mostly cloudy or overcast sky.

• allen_peak/frames/20260826_021531_00087.jpg

  Evening low-sun frame from the Allen Peak camera (ALERTCalifornia): dense conifer forest and layered ridges with grassy patches. Sun low in the west (5.7° altitude); waxing gibbous moon (96% illuminated) 4.6° up in the southeast. Colour image. Estimated conditions (automatic): partly cloudy sky.

• allen_peak/frames/20260828_032331_01561.jpg

  Evening nautical-twilight frame from the Allen Peak camera (ALERTCalifornia): dense conifer forest and layered ridges with grassy patches. Sun 8.2° below the horizon; full moon (100% illuminated) 7.9° up in the east, in partial lunar eclipse (umbral phase). Monochrome image (infrared/night mode).

• allen_peak/frames/20260826_094331_00311.jpg

  Repeated (stale) frame from the Allen Peak camera (ALERTCalifornia): dense conifer forest and layered ridges with grassy patches. Content identical to the earlier frame 20260826_094131_00310.jpg (first received 2026-08-26 02:40 local); the feed was not refreshed, so the image does not show conditions at request time (night, sun altitude -37.9°). Monochrome image (infrared/night mode).

• big_rock_ridge/frames/20260826_211331_00656.jpg

  Daytime frame from the Big Rock Ridge camera (ALERTCalifornia); PTZ position pan 164°, tilt -2°, zoom 12.7x (view not visually verified). Sun high in the southwest (58.9° altitude). Colour image. Estimated conditions (automatic): clear sky or few clouds.

Dataset creation

Curation rationale

The dataset supports research on outdoor scenes under changing natural light: estimating time of day and sun position from images, day/night and infrared transitions, cloud cover and glare, camera calibration in the wild, change detection and image captioning. A synchronised 72-hour window over many fixed and pan-tilt-zoom cameras gives three complete diurnal cycles per camera, a full-moon night with a partial lunar eclipse, and a wide range of terrain.

Source data

The frames were requested from the cameras' public "latest image" URLs by an automated capture script, one request per camera every 120 s, all cameras in lock-step, from 2026-08-25 23:23:31 UTC to 2026-08-28 23:21:31 UTC. The images are stored unchanged (no resizing, re-encoding or cropping); the logos, credits and pixelated privacy areas applied by the operators are part of the images. The capture log (request and completion times, byte counts, errors) is kept per frame in the record and per camera in cameras.json.

Annotation process

Captions and records combine five layers of evidence:

1. Scene and view descriptions (visual inspection). For every camera a 3x3 contact sheet of nine frames spread over the capture (morning, midday, afternoon, low sun, twilight, night, pre-dawn and a final-day midday frame) was inspected, plus one frame for every additional pan/tilt/zoom position the camera held for 20 frames or more. For every distinct framing ("view") a static description was written - terrain, land cover, notable objects, horizon type, approximate sky fraction - deliberately without lighting or weather words, because these change from frame to frame. Tree species are named only where they are reliably recognisable at webcam resolution (otherwise "conifer"). Each inspected frame also received its own caption (visual_check), and doubts were recorded rather than guessed (cameras.json, scene.reviewer_notes). In addition 1,010 randomly sampled colour frames with the sun above the horizon (four per camera; one for each of the two O'Fallon cameras, whose feeds were almost always stale) were labelled for cloud cover, fog, haze, smoke, precipitation, glare, lens contamination, obstruction and snow (visual_check.labels).
2. View of every other frame. Established, in order of strength, from: the frame itself being inspected; the pan/tilt/zoom values burned into the overlay (read by OCR) matching those of an inspected frame; stationary-segment and template matching on 16x9 thumbnails; and nearest-neighbour search on CLIP ViT-L/14 image embeddings against anchored frames of the same camera and illumination class (colour/monochrome x sun above/below 6°). A CLIP decision is accepted only if the best view's similarity is >= 0.93, at least 4 of the 5 nearest anchors agree and the margin to any other view is >= 0.02 (>= 0.96 when only one view has anchors in that class; 0.96 / 0.03 / 0.97 for colour frames with the sun below 6°). Stationary segments end wherever consecutive frames change abruptly in their thumbnails (correlation < 0.80) or CLIP embeddings (cosine < 0.90; < 0.80 across a colour/infrared switch). Frames that remained unassigned were clustered (CLIP cosine >= 0.93); for each cluster of >= 5 frames a mosaic of four members spread over time was inspected and matched to a view, defined as a new view, or rejected. 8 views that turned out to be one framing were merged after inspection, and cameras whose burned-in PTZ read-out stays constant while the framing changes (rasnow_peak_1) do not use the read-out. The method is recorded per frame in view.assignment.
3. Astronomy (computed). Sun and Moon positions, lunar phase and illuminated fraction were computed with Skyfield and the JPL DE421 ephemeris for the camera coordinates at the image time (topocentric apparent altitude, no atmospheric refraction, sea-level observer; an independent PyEphem computation agrees within 0.011°). Light phases follow the standard solar-altitude limits: day >= 6°, low sun -0.833° to 6°, civil, nautical and astronomical twilight down to -6°, -12° and -18°, night below. The partial lunar eclipse of 2026-08-28 is flagged in every frame taken while the Moon was above the horizon, using the published contact times of F. Espenak (EclipseWise.com): penumbral phase 01:23:29.0-07:01:59.8 UT, umbral (partial) phase 02:33:21.4-05:52:09.3 UT, greatest eclipse 04:12:52.0 UT, umbral magnitude 0.932 (an independent DE421 computation agrees within 35 s).
4. Pixel measurements (exact). Every JPEG was decoded and measured (see image.*). These give the image mode, an exposure class and exact detection of repeated images (byte-identical content). Daytime frames that correlate poorly (< 0.5) with their view's daytime template carry the flag low_similarity_to_view_template; the caption makes no claim about its cause (fog, glare, obstructions, but also large clouds).
5. Condition estimates (machine, validated). For colour frames with the sun above the horizon that are not repeated images, logistic-regression probes on the CLIP ViT-L/14 embeddings estimate cloud cover in three classes and sun glare / lens flare. They were trained on the 1,010 labelled frames plus 1,523 labelled contact-sheet frames and evaluated with fully nested, camera-grouped cross-validation (features, regularisation and thresholds chosen on inner folds only; reported numbers come from cameras never seen in training). An estimate is written only when the model's confidence clears the selected threshold, otherwise the value is null. An attribute is published only if it met the rule: cloud cover >= 0.92 accuracy, flags >= 0.90 precision and >= 0.25 recall. Fog, lens contamination, haze, smoke and precipitation did not meet it and are not estimated.

Who produced the annotations. The visual inspection in layers 1, 2 and 5 (scene and view descriptions, per-frame visual captions, condition labels, mosaic checks) and the audit below were carried out by an AI vision-language model (Anthropic's Claude), working frame by frame under the direction of the dataset author. No human crowd annotation was used. All other fields are computed deterministically or by the validated models described above.

Validation

View assignment. On held-out anchors of 74 multi-view cameras (anchors within +-60 min of the query excluded), the CLIP acceptance rules made no errors (one-sided 95 % upper bound on the error rate: 0.013 %) at 91.1 % coverage (22,440 of 24,623 queries).

Condition estimates (domain: 296,403 frames):

Attribute	Status	Nested camera-grouped CV (held-out cameras)	Values in this dataset
Cloud cover (3 classes)	published	94.5 % accuracy on emitted estimates (562 of 1,010 held-out frames, 56 % coverage); within one class 98.2 %	clear_or_few_clouds 87,937, partly_cloudy 16,103, mostly_cloudy_or_overcast 78,803; null 113,560
Sun glare or lens flare	published	true: precision 92.6 %, recall 74.8 % (122 emitted, 151 positives in 1,010); false: 98.3 % correct (71 % coverage)	true 31,923; false 226,316; null 38,164
Fog or low cloud	not published	precision 81.8 %, recall 22.5 % - below the publication rule	-
Lens contamination	not published	precision 92.5 %, recall 7.8 % - below the publication rule	-

Label agreement. A second, blind labelling pass over 120 frames from 30 cameras agreed with the first on cloud cover in 87.5 % of cases with five classes (weighted kappa 0.91) and 90.8 % with three classes (kappa 0.86); kappa was 0.93 for sun glare, 1.00 for fog, 0.70 for lens contamination, 0.56 for heavy haze and 0.49 for precipitation (attributes with kappa < 0.6 are not estimated). Both passes were made by the same model, so this measures consistency rather than agreement between independent humans.

Audit. A separate verification pass re-checked 74 randomly sampled, previously uninspected frames (stratified over view-assignment methods, estimate classes, night and unassigned frames) against their captions, judging only the pixels. View statements: 69 of 71 confirmed, 0 wrong, 2 not verifiable (too dark or obscured). Image mode and light phase: 74/74 and 74/74 correct. Cloud-cover estimates agreed with the independent label in 33 of 34 cases; glare flags: 9/9 "true" and 35/37 "false" correct. Two problems found by the audit were fixed before release (a runway described as a dirt airstrip; a caption sentence triggered by large clouds, now removed).

Statistics

Light phase at image time

Light phase	Frames
day	274,364 (50.03 %)
low_sun	26,741 (4.88 %)
civil_twilight	20,622 (3.76 %)
nautical_twilight	24,615 (4.49 %)
astronomical_twilight	25,621 (4.67 %)
night	176,413 (32.17 %)

How the view of each frame was established

Method	Meaning	Frames
overlay_ptz_match	burned-in pan/tilt/zoom equals that of an inspected frame	341,889 (62.35 %)
image_segment_with_anchor	same stationary image segment as an anchored frame (thumbnail and CLIP continuity)	189,920 (34.63 %)
image_template_match	thumbnail matches a verified view template (corr >= 0.90, margin >= 0.05); not contradicted by CLIP	10,318 (1.88 %)
visually_verified_frame	the frame itself was inspected	2,482 (0.45 %)
image_embedding_match	CLIP ViT-L/14 nearest anchored frames of the same illumination class agree (validated, see below)	1,312 (0.24 %)
unassigned	view could not be established; the caption says "view not visually verified"	814 (0.15 %)
single_view_camera	fixed (non-PTZ) webcam with one verified framing; frame not otherwise anchored	699 (0.13 %)
temporal_interpolation	no PTZ read-out; the previous and next assigned frames show the same view	439 (0.08 %)
image_embedding_cluster	member of a cluster of near-identical frames (CLIP cosine >= 0.93) whose sample was inspected	385 (0.07 %)
overlay_ptz_near_match	PTZ within 3 deg pan / 2 deg tilt / 15 % zoom of an inspected frame	118 (0.02 %)

Image mode

Mode	Frames
colour	344,966 (62.91 %)
monochrome (infrared / night mode)	203,372 (37.09 %)
almost entirely black	38 (0.01 %)

• Image time source: capture_request_time 447,268, overlay_clock_ocr 101,108.
• Frames taken during the lunar eclipse with the Moon up: partial_umbral 24,619, penumbral 9,523.
• Repeated (stale) images: 10,310 frames in 253 cameras (largest: o'fallon_weather_(west) 2,158, o'fallon_weather_(south) 2,157, los_pinos_south 1,352, grass_mountain_1 996, buck_rock_2 585).
• Flags: identical_to_previous_frame 10,310, repeated_image 10,310, overlay_clock_repeated 1,871, low_similarity_to_view_template 850.
• Individually inspected frames: contact_sheet_tile 2,286, condition_label_sample 1,008, unassigned_cluster_check 152, ptz_state_check 53.
• Cameras with operator privacy masking (pixelated or grey areas): 26.

Uses

Intended: research and teaching on illumination and time-of-day estimation, sun and moon geometry, day/night and infrared transitions, cloud and glare recognition, camera calibration, change detection, captioning and vision-language grounding, and studies of wildfire-camera imagery.

Out of scope: identifying, tracking or monitoring people or private property; any use that conflicts with the camera operators' terms; commercial use (the annotations are licensed for non-commercial use only).

Personal and sensitive information

The cameras show landscapes from mountain tops, towers and buildings; some frames show houses, roads, vehicles, airfields and helipads, and people may occasionally be visible at very low resolution. 26 cameras carry privacy masks applied by the operator; they were left untouched. No personal data were collected, and no attempt was made to detect or identify people. Camera coordinates are the public site locations published by the operators.

Bias, risks and limitations

• Geographic and seasonal coverage is narrow: late-summer California landscapes dominate (247 of 254 cameras), with dry grassland, chaparral and conifer forest; three days of one season.
• Only 3,499 frames were inspected individually. All other captions combine a verified view description with computed and measured facts and clearly labelled machine estimates; they do not describe transient content (smoke, vehicles, animals, precipitation) of uninspected frames.
• Condition estimates are statistical: on held-out cameras the published cloud-cover estimates were right about 94 % of the time and the glare flags had about 93 % precision. null means "not confident", not "absent". No estimates are made at night, in twilight or for monochrome frames.
• View descriptions are static; at night and in infrared frames they describe the daytime appearance of the framing (colours named in a description are not visible then).
• 814 frames could not be tied to a verified view; their captions say so.
• Times in file names are scheduled request times; the exposure can precede them by up to about a minute (time.overlay_exposure_utc, time.image_time_utc).
• The annotations were produced by an AI model and automated methods; despite the validation above, individual errors remain possible. Please report them in the Community tab.

Licensing

• Annotations - captions, records, metadata/frames.parquet, cameras.json, captions.json, the schema, this card and the script - are licensed under the Creative Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0).
• Images are not covered by that licence. Copyright remains with the camera operators and image providers: ALERTCalifornia (UC San Diego) and the organisations credited in the image overlays or named in cameras.json (e.g. PG&E, CAL FIRE, SDG&E and the partner networks in site_partner_tag), the Albuquerque Soaring Club, NVWX (Elko), the O'Fallon weather station and Ajijic Weather. They are included for non-commercial research and education; users must check that their use complies with the operators' terms.
• Rights holders who want images removed can open a discussion in the Community tab; requests will be handled promptly.

See LICENSE for the full terms.

Citation

@misc{webcam254_2026,
  title        = {WebCam254: A Captioned 72-Hour Time-Lapse Dataset from 254 Outdoor Webcams},
  author       = {Alexander Ezharjan},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Ezharjan/WebCam254}}
}

Computation and provenance

• CLIP embeddings: openai/clip-vit-large-patch14, fp16, two 224 px square crops (left and right) per frame, computed on NVIDIA GeForce RTX 5060 Laptop GPU with torch 2.13.0.dev20260516+cu130 / transformers 5.5.3 (CUDA 13.0); 548,376 frames embedded, 0 undecodable.
• Probe cloud3: features = embedding and embedding minus the camera-view mean, contact-sheet frames used in training, C = 10, confidence threshold 0.88.
• Probe glare: features = embedding minus the camera-view mean, contact-sheet frames used in training, C = 10, threshold true >= 0.74, false <= 0.27.
• Astronomy: Skyfield with JPL DE421. Pixel measurements: Pillow and NumPy over every JPEG. Overlay text: OCR.
• No external weather data were used; every derived value is reproducible from the images and the records.

Version

1.0 (2026-09-27): first release; record schema 1.2.0.