Measure Water Change From Masks

Measure how the visible water area changed between two images of one fixed camera, from two water masks a person accepted. Evidence for human review only: not flood detection, not water height, not flow speed, no alert. The full guide is docs/learning/water-change-measurement.md; this page keeps the same metrics and limits.

What is measured

For an ordered pair (earlier, later) inside the fixed watched area (the ROI the segmentation used):

• Water fraction of each image: water pixels ÷ ROI pixels.
• Change in percentage points: later − earlier, signed.
• Newly wet and no longer wet fractions: pixels that changed, ÷ ROI pixels. Equal totals in a different shape are therefore still visible.
• Change per hour (optional): an image-space area rate, not a level rise or speed.

One denominator (ROI pixels). Several detections in one result are combined with OR. Only pixels inside the ROI are measured; a mask with water outside it is refused. Coordinates are original image pixels, right and bottom edges exclusive. Brightness is never read.

Unavailable, never zero

Missing, unreviewed, rejected, needs-correction or non-water masks; changed images; a different camera, image size or watched area (including a mask cut from a different watched area than the run froze); reversed or equal timestamps; or unconfirmed framing give an unavailable or invalid record with reason codes. An accepted "no match" is not a verified empty mask. A named person confirms the fixed framing; there is no automatic camera alignment.

Saved evidence

Each result is frozen once in outputs/water-change-pairs/ with image and mask hashes, run and result ids, review decision, configuration, calculation version and frozen gauge context. Original runs are never modified. The adapter water_change_mask_v1 is off by default.

The pilot

10–15 diverse pairs from one camera (rising, falling, stable-high, equal area different shape, difficult). Write the criteria first; two reviewers judge blind to machine results and gauge (they can label on the focused review page and export their blind labels with the judgments command); then measure, view the contact sheet and report; show frozen gauge context last; a person records proceed, revise or stop. Errors and the unavailable rate are both reported, held-out pairs are scored only on request and logged, and unknown-for-everything cannot pass. Commands are in scripts/water_change_pilot.py.

The Review chart: pixel change or water coverage

The second Review chart has a Measurement switch with two choices, each explained under the switch. Pixel appearance change (the default and original region change score) is how different the watched area looks from the baseline, from pixel brightness. Segmentation: water coverage is, for each image, the share of the watched area the segmentation marks as water, in percent, so the line shows how visible water changes from image to image (the picture, not depth or flow). Filled points use a mask a reviewer accepted; hollow points use an unreviewed draft; an image with no usable mask has no point, never zero. Masks must share one image size and watched area, camera movement is not detected, and the two numbers are never combined. The choice is remembered in your browser only.

Compare any two images in Review

Optional, in Review: Compare with another image… next to the comparison view switch. Pick any other saved image of the same camera (across sequences and runs, with thumbnails, exact times and the mask state; the same file name in two sequences counts as two images). The existing Side by side and Overlay views then show the two images earlier then later, whichever you picked first.

With accepted water masks on both and your confirmation that the view did not move, it shows Water coverage change: coverage at each end, the signed change in percentage points, newly wet and no longer wet areas, an image-space rate, and a picture of where water arrived and left. If a number is not possible (missing, unreviewed or rejected mask, different watched area or size, equal times, unconfirmed view) you still see the images with a plain reason. It is separate from the run's Pixel appearance change score and never combined with it. Viewing writes nothing and starts no segmentation; Save this comparison is a separate, explicit button that never overwrites an earlier result. Closing returns you to the same image, filters and scroll position.

Limits

Image-space area only; hosted masks with no editor; person-confirmed framing; a few pairs from one camera cannot show accuracy. No training, alerts, risk-engine wiring, public upload, extra baseline lines or gauge-height regression are part of this.