Find your next learning step
Pixels become operational events only after a connectivity rule, a size threshold and an object-matching decision. Those choices can erase small incidents, merge neighbours or split one event into several alerts even when the pixel map is unchanged.
Open the chapter contents · Remote Sensing root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Build alert objects
Route 2 · Work the event ledger
Route 3 · Test thresholds and boundaries
Route 4 · Check evidence and limits
Route 5 · Ask and continue
CHAPTER 1 OF 12 · BUILD ALERT OBJECTS
1. A cluster is made by a rule
A thresholded alert image becomes an object only after a neighbourhood rule joins pixels. Four-connectivity and eight-connectivity can produce different components from the same diagonal pattern. Morphological closing can join gaps; opening can remove narrow pieces.
State the pixel threshold, connectivity, gap-filling operations, minimum area and map resolution. The earlier guide on spatial alert clustering explains why pixel and incident denominators differ. This continuation follows the geometry of the incidents.
CHAPTER 2 OF 12 · BUILD ALERT OBJECTS
2. Size thresholds change the event population
A minimum cluster size removes small components. It can suppress isolated noise, but it can also remove real small incidents or fragments created by cloud, shadow and missing data.
Report performance across a range of size thresholds. A threshold selected on the evaluation set can make the final accuracy optimistic unless the selection step is reproduced on new data.
CHAPTER 3 OF 12 · WORK THE EVENT LEDGER
3. Work an object-matching ledger
Suppose the reference contains incidents A, B and C with areas 80, 60 and 50 pixels. The alert map contains M1, M2 and M3 with areas 110, 38 and 18. M1 overlaps A by 70 pixels and B by 30; M2 overlaps C by 30; M3 overlaps C by 12.
Using intersection over union, IoU(M1,A) = 70/(110 + 80 − 70) = 0.58. IoU(M1,B) = 30/(110 + 60 − 30) = 0.21. IoU(M2,C) = 30/(38 + 50 − 30) = 0.52, while IoU(M3,C) = 12/(18 + 50 − 12) = 0.21.
| Mapped object | Main overlaps | Best IoU | Geometric issue |
|---|---|---|---|
| M1 | A and B | 0.58 with A | two references merged |
| M2 | C | 0.52 | main part of a split |
| M3 | C | 0.21 | small split fragment |
CHAPTER 4 OF 12 · WORK THE EVENT LEDGER
4. One-to-one matching gives an event count
Set an IoU threshold of 0.50 and require one-to-one matches. M1 matches A and M2 matches C. B is missed, and M3 is an unmatched alert. Event precision is 2/3 and event recall is 2/3.
That summary is useful but incomplete. M1 also overlaps B, so a merge helped it earn one match while hiding another reference. M2 and M3 split C. Retain split and merge counts beside precision and recall.
CHAPTER 5 OF 12 · WORK THE EVENT LEDGER
5. A size threshold can improve a score selectively
Apply a 20-pixel minimum. M3 disappears, so event precision becomes 2/2 = 1.00 while recall remains 2/3. The score improved because a small unmatched fragment was removed.
The decision may be reasonable, but it should be visible. If genuine small incidents also fall below 20 pixels, the same rule changes the target population and may reduce practical sensitivity.
CHAPTER 6 OF 12 · TEST THRESHOLDS AND BOUNDARIES
6. Changed context: allow a boundary tolerance
Dilate each reference by one pixel before assessing whether mapped boundaries are acceptably located, or use a distance-weighted boundary measure. IoU values can rise for narrow or slightly displaced objects even though their centres and event identities are unchanged.
Recompute every match with the same tolerance. Do not tune a separate tolerance for each object. Report both strict overlap and tolerance-aware geometry so the reader can see what the sensor resolution justifies.
CHAPTER 7 OF 12 · TEST THRESHOLDS AND BOUNDARIES
7. Split and merge errors need their own ledger
Define a split when two or more mapped objects substantially overlap one reference, and a merge when one mapped object substantially overlaps two or more references. Specify “substantially” through overlap area, proportion or distance.
An incident response team may prefer one merged alert to ten fragments, while ecological mapping may need separate patches. The loss function must follow the use case rather than a generic object score.
CHAPTER 8 OF 12 · TEST THRESHOLDS AND BOUNDARIES
8. Boundaries may be genuinely indeterminate
Coastlines, burn edges, flood margins and vegetation transitions can move during acquisition or fade gradually across pixels. A single crisp reference polygon may overstate certainty.
Use repeated delineations, confidence bands, soft boundaries or centre-weighted metrics when justified. Keep acquisition times and positional uncertainty with the reference data.
CHAPTER 9 OF 12 · CHECK EVIDENCE AND LIMITS
9. Limits and misconceptions
Object accuracy is not automatically superior to pixel accuracy. It answers a different question and introduces connectivity, segmentation and matching choices. A high event precision can coexist with poor area estimation.
The worked example ignores geolocation uncertainty and gives exact overlaps. Real objects can be nested, partially observed or temporally mismatched. The ledger is a transparent starting point, not a universal benchmark.
CHAPTER 10 OF 12 · CHECK EVIDENCE AND LIMITS
10. A practical evidence check
Freeze thresholds before final evaluation. Save the probability surface, binary mask, component labels and reference versions so every event can be reconstructed.
Report pixel confusion, object precision and recall, area error, boundary distance, split count, merge count and performance by reference size. Inspect cases near the size and IoU thresholds rather than only the average.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Does removing tiny clusters always remove noise?
No. It removes whatever falls below the threshold, including genuine small events and fragments caused by missing observations.
Why not match every mapped object to its nearest reference?
Nearest distance alone can allow several alerts to claim one incident or can match objects with little meaningful overlap. The matching rule must enforce the intended event logic.
Should uncertain boundaries be ignored?
No. Model their tolerance explicitly and show how conclusions change. Quietly enlarging references can inflate apparent accuracy.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Continue through verified routes
Start with the Remote Sensing root, Gold Standard of Education and How X Works. Continue through Rare Change and False Alarms, Coregistration and False Boundary Change and the spatial-clustering guide linked above.
For specialist methods, use the Remote Sensing paper on good practices for object-based accuracy assessment and the study of accuracy measures for object extraction. The split–merge ledger is an original teaching example.
Top · Contents · Previous · How X Works