A satellite revisiting a location does not mean a monitoring team receives a usable answer at that moment. An optical image may be obscured by cloud. A clear image may have unsuitable resolution for the feature of interest. A suitable observation may arrive after the operational decision has already been made.

This distinction matters when a business buys or builds a monitoring service. “New imagery every few days” sounds like a service commitment, but it describes only one part of the chain. This article follows a fictional land-change monitoring task to explain how to turn satellite availability into a realistic information service, including observation quality, processing, validation and the gaps that remain.

Start with the decision on the ground

Suppose a land manager wants to identify possible new clearance near a set of infrastructure corridors. A reviewer will inspect flagged areas and decide whether a ground investigation is warranted. The service is a screening tool; it does not establish who caused a change or prove that an activity breached a rule.

Define the smallest change that matters, the area to monitor, the timing of the decision and the consequence of a missed or false alert. A narrow access track, a large cleared field and a seasonal vegetation change are different detection problems. The phrase “monitor land cover” does not specify which one the system must solve.

Also define the fallback. If no usable observation is available during an important period, does the team wait, request another data source or use a ground inspection? An honest “no current evidence” status helps the organisation act. A dashboard that keeps showing an old green map can create false reassurance.

Myth: repeat visits mean equally frequent answers

Acquisition opportunity is the chance for a sensor to observe the location. Usable-observation interval is the gap between observations that satisfy the task’s quality requirements. Delivery latency is the time between acquisition and product availability. Decision latency includes the processing and review needed before someone acts.

These intervals are related, but they are not interchangeable. A service can improve its processing speed without changing cloud conditions. It can acquire more imagery without improving its ability to resolve a small target. A useful design identifies which interval is currently limiting the decision.

Illustrative observation timeline, not a satellite schedule: frequent acquisition opportunities can still leave a long gap in usable evidence.Day 0UsableDay 5CloudDay 10ShadowDay 15Usable15 days between usable observations
Illustrative observation timeline, not a satellite schedule: frequent acquisition opportunities can still leave a long gap in usable evidence.

For an optical monitoring task, review historical observations over representative seasons. Count observations that were actually useful within the area of interest. A theoretical repeat interval is not a substitute for that analysis, and a year with favourable conditions does not guarantee the next year will be similar.

Myth: a low scene-cloud percentage means the target is clear

A scene-wide cloud percentage can help filter a large catalogue. It does not tell you whether a specific corridor is clear. A tile with little cloud overall can still have cloud directly over the asset. Conversely, a cloudy tile may contain a clear area that is useful for the task.

The Copernicus Sentinel Hub documentation explicitly notes that its maximum-cloud-coverage filter uses tile-level estimates and may not represent the chosen area of interest. For the monitoring service, use appropriate pixel-level quality information and inspect how it behaves over the target area, rather than treating a catalogue threshold as a guarantee.

Sentinel-2 Level-2A processing provides surface reflectance and scene-classification information. These are useful inputs to a quality workflow, not a complete validation of the final change alert. Thin cloud, shadows, snow, mixed pixels and classification errors can still influence interpretation. Document which observations are excluded and why.

Myth: a colourful difference map proves a change

A change detector compares measurements that may differ for reasons unrelated to land change. Seasonal growth, illumination, viewing conditions, atmospheric correction, alignment and processing choices can affect the output. Begin with a baseline period that fits the phenomenon, not simply the oldest image available.

For the fictional clearance task, comparing a leaf-on scene with a leaf-off scene could generate widespread differences that are unhelpful to the reviewer. A more appropriate baseline may use comparable seasonal conditions or a time-series method that models normal variation. The choice should follow the task and be tested against examples.

Preserve the source product identifiers, acquisition dates, masks, processing version and configuration for each finding. If a later version changes an alert, the team should be able to determine whether the source data changed, the quality rule changed or the analytical method changed.

Reality: follow the evidence behind one alert

Imagine a new observation contains a candidate clearing beside one corridor. The processing pipeline flags a difference from the baseline. Before presenting it as a finding, the system checks that both dates have sufficient valid coverage over the candidate and that the difference is not simply an alignment edge or a cloud shadow.

The reviewer receives the two source views, the acquisition dates, the highlighted candidate and the relevant quality information. The interface calls it a “candidate change for review”, not a confirmed incident. If the reviewer cannot resolve it from imagery, the outcome remains uncertain and follows the agreed investigation process.

Information needed to interpret one candidate change
RecordWhy it matters
Baseline and new acquisition datesDefine the period during which the change may have occurred
Valid coverage over the candidateSeparate an observed difference from missing information
Source and processing identifiersAllow the result to be reproduced and investigated
Reviewer outcome and reasonDistinguish a confirmed change, false alert and unresolved case
Next operational actionConnect the analysis to a responsible person and decision

The same design must handle a quieter outcome: no alert because the corridor was not observed clearly. That result should not be counted as evidence that no change occurred. Display the age and coverage of usable evidence alongside the alert status.

Synthetic before/after map for explanation only. A candidate change and an obscured area are different evidence states; neither map is real satellite imagery.Earlier usable observationLater candidate observationGold outline: candidate change Grey: obscured area
Synthetic before/after map for explanation only. A candidate change and an obscured area are different evidence states; neither map is real satellite imagery.

Evidence: validate the errors that affect the decision

Use independently checked examples, separated from those used to tune the method. Include real changes, unchanged areas and difficult conditions that resemble change. Where reliable ground reference is unavailable, state that limitation instead of treating agreement between two processing methods as proof.

Measure both missed changes and unnecessary alerts. A system that flags nearly everything may find many real changes while overwhelming the review team. A strict detector may produce attractive-looking alerts while missing smaller events that matter. Evaluate the workload created for reviewers as well as the detection results.

Break results down by relevant conditions: change size, season, land-cover type and quality of available observations. A single overall accuracy number can conceal a weak category that dominates the organisation’s risk. Keep the evaluation geographically and temporally relevant to the intended deployment.

What to do instead: specify an information service

A practical monitoring commitment describes its area, detectable features, quality rules, delivery process, review responsibilities and missing-data response. It distinguishes a target from a guarantee and names the assumptions behind each. Historical observation analysis can inform that commitment, but it cannot remove future uncertainty.

Before choosing a satellite product, write a one-page decision brief and examine a representative sample of observations for the actual area. The question is not only how often a sensor can return. It is how often the organisation can obtain enough trustworthy information to take the next useful action.

Technical references

The corridor scenario and timeline are illustrative, not measurements or a claim of operational detection performance.