Does collocation require exactly identical coordinates and times?

Usually practical comparisons use defined tolerances or other matching procedures because exact correspondence is uncommon.

It can support a particular comparison, but its suitability depends on the physical quantity, variability, and sampling design; it can also conceal meaningful variation.

Is the reference always perfect?

No. Reference observations have their own uncertainties and may represent a different spatial or temporal scale.

AI Glossary / Data science and analytics / Intermediate

Collocation (remote sensing)

Definition

Collocation in remote sensing is the matching of observations from different sources so that they can be compared as measurements of sufficiently corresponding places, times, and physical conditions. A common application pairs satellite observations with ground-based or airborne measurements. Because instruments rarely observe exactly the same volume at exactly the same time, the matching criteria and remaining mismatch are central to interpretation.

What it means

A satellite pixel may summarize an area, while a ground instrument observes a much smaller location or samples conditions over time. Even if their recorded positions are close, their measurements need not represent identical material. Collocation creates a defensible comparison under stated rules; it does not make the observations physically identical. The term here concerns measurement matching, not linguistic word combinations or placing computing equipment in a shared facility.

How it works

Identify the physical quantity being compared, the instruments' spatial coverage, timing, and relevant vertical or spectral characteristics. Apply quality filters and choose matching criteria appropriate to the phenomenon and purpose. Form matched observations, using documented averaging or interpolation where justified. Retain distances, time offsets, and other relevant matching metadata. Assess whether differences between paired values reflect instrument behavior, retrieval methods, natural variability, or the matching procedure itself. Sensitivity checks with alternative reasonable criteria can help reveal dependence on the chosen comparison rules.

Three practical examples

Satellite validation: researchers compare satellite aerosol estimates with suitable ground-based observations while accounting for differences in spatial and temporal sampling.
Atmospheric profiles: a team pairs satellite-derived profiles with radiosonde observations, examining differences in observation locations and times.
AI remote sensing: developers match satellite inputs with reference observations for evaluation, documenting how labels correspond to the satellite footprint and acquisition time.

Think of it this way

Think of comparing two photographs of a busy market taken from different viewpoints a few minutes apart. Differences may come from the cameras, but people and lighting may also have changed. The analogy explains why proximity alone does not guarantee comparability. Scientific collocation additionally requires attention to what physical quantity each instrument measures and how it samples that quantity.

Worked example

Suppose a fictional study defines a teaching-only rule: a reference observation qualifies if it is within 5 kilometers and 15 minutes of a satellite observation. Candidate A is 3 kilometers away and 10 minutes earlier, so it meets both criteria. Candidate B is 2 kilometers away but 25 minutes later, so it fails the time criterion. Candidate C is 7 kilometers away and 5 minutes earlier, so it fails the distance criterion.
Only A qualifies under this rule, but qualification does not prove it represents the entire satellite footprint. A rapidly changing local condition could still create a substantial difference. If the researchers widen the time limit to 30 minutes, B becomes eligible; that increases potential matches while changing the sampling mismatch. These invented limits are arithmetic examples, not recommended thresholds for any real sensor or application.

Illustrative case study

In a fictional environmental analytics startup, developers evaluate an AI model against field observations. They initially match each image to the nearest station record without considering acquisition times. A review finds that some pairs span a substantial change in conditions. The team introduces an explicitly justified time criterion, preserves the original offsets, and reports how many observations remain eligible. It also checks whether the remaining sample overrepresents stable conditions. The revised comparison is more transparent, but the scenario does not claim that stricter filtering automatically produces a representative or error-free evaluation.

Common misconceptions and mistakes

A nearby reference is not automatically ground truth for an entire pixel. Smaller matching windows are not always unconditionally better: they may reduce mismatch while leaving fewer observations or increasing sensitivity to measurement noise. A difference between collocated measurements cannot automatically be assigned to the satellite instrument. Matching also differs from data fusion: creating paired observations for comparison does not itself combine them into a new estimate. Finally, a single set of distance and time thresholds should not be treated as universal across rapidly and slowly varying phenomena.

Business use

Businesses using remote-sensing products should ask how reference data were matched before relying on published accuracy claims. The relevant questions include observation scale, time offsets, quality exclusions, and the environments represented by the matched sample. For AI projects, a poorly matched reference can make a useful model appear inaccurate or make a misleading evaluation look strong. Preserve matching metadata alongside results so later reviewers can distinguish model changes from changes in the evaluation sample or collocation procedure.

Check your understanding

Two measurements meet a distance threshold, but one was taken before a rapidly changing event and the other afterward. Can their difference automatically be treated as instrument error?

Answer: No. Temporal mismatch and real environmental change may contribute. The comparison needs suitable time matching and uncertainty interpretation.

Related terms

Data fusion
remote sensing
sensor validation
spatial resolution
temporal resolution
footprint
representativeness error
measurement uncertainty

Learn next

sketch the space and time covered by two observing systems. Identify which aspects overlap and which do not. Then vary an illustrative matching tolerance and track both the number of matched pairs and the kinds of conditions retained, rather than optimizing the pair count alone.

Sources

Sources: Virtanen and colleagues' original study explains collocation mismatch in satellite aerosol validation: https://amt.copernicus.org/articles/11/925/2018/amt-11-925-2018.html . Gilpin, Rieckh, and Anthes examine representativeness and sampling errors in satellite–radiosonde comparisons: https://amt.copernicus.org/articles/11/2567/2018/ . Abstracts and relevant introduction passages were read. Their study-specific numerical findings and matching distances are not generalized here; all worked-example thresholds are deliberately fictional.

Previous: Collinearity Next: Cologne phonetics