Hard-Case Validation Examples

This page lists confusing cases OpenFloodAI should handle carefully.

Simple meaning: these are the situations where the system should say "I cannot safely judge this" instead of acting confident.

These examples are for validation and review only. They do not prove flood detection accuracy, send alerts, upload files, or create public warnings.

The matching machine-readable fixture is:

data/sites/example-site/expected-behavior/hard-cases.jsonl

Real hard-case sample metadata should use:

data/sites/<site-name>/expected-behavior/real-hard-case-samples.jsonl

Simple meaning: hard-cases.jsonl says what the system should do for confusing case types. real-hard-case-samples.jsonl tracks real examples when we find them.

Expected Behavior

Case Input Quality Validation Result Simple Reason
Missing video UNKNOWN cannot_compare There is no video to process.
Empty video UNKNOWN cannot_compare There are no usable frames.
Unreadable video UNKNOWN cannot_compare The file cannot be read as video.
Camera offline UNKNOWN cannot_compare The camera-style input is unavailable.
Heavy glare DEGRADED cannot_compare Bright glare can hide the watched water area.
Rain or noisy image DEGRADED cannot_compare Rain or noise may look like water movement.
Night or dark frame DEGRADED cannot_compare The watched area is too dark to judge safely.
Camera shake DEGRADED cannot_compare The image may move because the camera moved.
Blocked view DEGRADED cannot_compare Something blocks the watched area.
Compression artifacts DEGRADED cannot_compare Video artifacts can create fake pixel changes.

Plain Example

Case: glare-001
Input quality: DEGRADED
Validation result: cannot_compare
Reason: Bright glare makes the watched water area unclear.

This does not mean the river is safe. It means the system does not have clear enough evidence.

Review Rule

Hard cases should stay visible in reports.

  • cannot_compare is not success.
  • UNKNOWN is not normal.
  • DEGRADED means the evidence may be weak or confusing.
  • A person should review the video or fix the camera/input before using the result.

Simple example: if the camera is shaking, many pixels may change. That does not prove the water changed. The report should explain the uncertainty.

For real-world samples, see the Real Hard-Case Sample Workflow.