Find your next learning step
A sequence of valid detections is not automatically one continuing event. Across revisit gaps, an object may move, split, merge, disappear behind cloud or be replaced, so identity must be assigned with an explicit rule and uncertainty.
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 · Define event identity
Route 2 · Work the revisit ledger
Route 3 · Test persistence and gaps
Route 4 · Check evidence and limits
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE EVENT IDENTITY
1. Identity is a claim across time
An object detection answers “what region was found in this observation?” Persistence asks a harder question: “is a region at a later time the same physical event?” Similar position, shape or spectrum can support that claim, but none alone defines identity.
State the event model. A flood patch may grow continuously, a fire perimeter may split, a vessel may move, and a construction site may disappear because work ended. The identity rule must suit the process rather than forcing every subject into a rigid box track.
CHAPTER 2 OF 12 · DEFINE EVENT IDENTITY
2. Revisit gaps hide transitions
Orbital imagery samples selected times, often with additional cloud or quality gaps. A feature present on Monday and Friday need not have persisted continuously; it could have disappeared and returned. A feature absent on Wednesday may have been obscured rather than gone.
Distinguish observation absence from event absence. The revisit-gaps and missed-events guide provides the corresponding sampling argument.
| Observation state | Event state | Safe wording |
|---|---|---|
| detected | present at acquisition | observed present |
| clear non-detection | possibly absent | not detected under stated sensitivity |
| cloud or invalid | unknown | unobserved |
CHAPTER 3 OF 12 · WORK THE REVISIT LEDGER
3. Work one event ledger
At Day 0, two alert objects A and B are mapped. At Day 3, cloud covers A while B is detected. At Day 6, one larger object C overlaps 60% of A’s former area and 30% of B’s. A one-to-one assignment rule links C to A because it has the larger overlap, leaving B’s track unmatched.
The ledger now contains one continued track, one fragment ending on Day 3 and an unresolved possibility that A and B merged. A single label cannot preserve both one-to-one identity and the merge history.
| Day | Usable observation | Detections | Track decision |
|---|---|---|---|
| 0 | yes | A, B | start two tracks |
| 3 | partial cloud | B | keep A dormant; update B |
| 6 | yes | C | assign C→A; flag possible merge |
CHAPTER 4 OF 12 · WORK THE REVISIT LEDGER
4. Fragmentation is not the same as error count
A track fragment occurs when one ground-truth trajectory is interrupted and later represented by another estimated segment. Identity switches count changes from one assigned identity to another. Missed detections, fragments and switches describe related but different failures.
Report detection precision and recall alongside identity measures. A tracker can preserve identities by refusing difficult assignments, producing fewer switches but more missed observations.
CHAPTER 5 OF 12 · WORK THE REVISIT LEDGER
5. Use gates before scores
Candidate matching normally begins with admissibility gates: maximum displacement, plausible area change, spectral compatibility, topology or land context. A cost can then combine distance, overlap and appearance for admissible pairs.
Do not let a weighted score rescue a physically impossible match. Register images to a common reference and propagate positional uncertainty into the gate. The cluster-threshold and split–merge guide handles the object-construction step before tracking.
CHAPTER 6 OF 12 · TEST PERSISTENCE AND GAPS
6. Changed context: double the gap
Suppose the maximum plausible displacement is 500 m per three-day revisit. With a six-day gap, a linear bound might become 1 km, admitting many more candidates. If behaviour can change direction, the uncertainty region grows in two dimensions rather than along one line.
A looser gate reduces premature track death but raises false-link risk. The correct change is not automatically “double every threshold”; use a motion or process model whose uncertainty expands with elapsed time.
CHAPTER 7 OF 12 · TEST PERSISTENCE AND GAPS
7. Splits and merges need lineage
When one region becomes two, a strict one-to-one tracker must choose one child or end the parent. Event analysis often needs a lineage graph instead: one parent can connect to several descendants, and several parents can connect to a merged region.
Keep the operational track ID separate from the physical-event family ID. This prevents a necessary database convention from being mistaken for a scientific claim that the event stayed indivisible.
CHAPTER 8 OF 12 · TEST PERSISTENCE AND GAPS
8. Persistence can be probabilistic
Rather than declaring a dormant track alive for exactly N days, maintain a probability or confidence that reflects elapsed time, missing-observation quality and process dynamics. Clear non-detections should usually reduce confidence more than cloud-covered acquisitions.
Calibrate any probability against labelled sequences. A score used only for ranking is not automatically a calibrated chance of persistence.
CHAPTER 9 OF 12 · CHECK EVIDENCE AND LIMITS
9. Limits and misconceptions
Satellite video offers dense frames over a short viewing interval; ordinary orbital revisit series offer sparse acquisitions over longer gaps. Methods and metrics from video tracking are informative, but they cannot be transferred unchanged to intermittent event maps.
High overlap does not prove identity, and low overlap does not disprove it for fast or expanding events. Registration error, sensor change, resolution, illumination and classifier drift can all alter matching evidence.
CHAPTER 10 OF 12 · CHECK EVIDENCE AND LIMITS
10. A practical evidence check
Create labelled sequences with explicit occlusion, split, merge, disappearance and reappearance cases. Freeze the detection and matching thresholds before evaluation. Report detection metrics, identity F1, switches, fragments, mostly tracked events and lineage accuracy where relevant.
Stratify results by gap length, cloud condition, object size and motion. Audit a sample of confident matches and rejected candidates. Keep the original masks, uncertainty layers and assignment costs so an event history can be reconstructed.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Is every reappearance the same event?
No. It is a hypothesis supported by spatial, temporal and thematic evidence, with alternatives retained when the data cannot decide.
Should a cloud gap end a track?
Not automatically. It creates missing observation, while the allowed dormancy should follow the event model and calibrated evidence.
Can one ID represent a merge?
It can satisfy an operational convention, but a lineage representation is needed to preserve both parent histories.
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 Temporal Frame Ties, Spatial Alert Clustering and the revisit and cluster guides linked above.
Primary research routes include Ciaparrone and colleagues’ tracking of pedestrians and vehicles in aerial imagery, HRTracker’s satellite-video data-association study and OITrack’s trajectory-completion study. The revisit ledger is an original teaching example and deliberately distinguishes sparse revisits from video frames.
Top · Contents · Previous · How X Works