A workflow that succeeds on a small study area may encounter different constraints when the number of observations, locations, or users increases. Scaling requires attention to processing capacity, data organisation, reference coverage, and the effort needed to review results. This perspective looks beyond whether a calculation can run on a larger dataset. It asks whether the complete service can preserve its meaning, quality, and delivery commitments as its scope expands. The objective is a design that identifies bottlenecks early and increases throughput without silently weakening the evidence available to users.
This working guide considers designing ground validation for satellite products in onion production beds. The immediate decision is to review beds with uneven establishment. Useful reference evidence includes planting dates, bed cover and crop-stage observations. The case is an illustrative planning context, with assumptions to check against the actual sensor, landscape and project requirements.
The setting: Onion production beds
Onion fields can retain a strong soil contribution in optical imagery because of their canopy form and the arrangement of production beds. The proportion of exposed soil, the stage of the crop and the condition of neighbouring surfaces all matter to interpretation. A useful monitoring design begins by identifying whether it needs a field-level development indicator or more detailed bed-level information. The image data must genuinely support the scale at which conclusions are presented.
Maintain planting and management records alongside the image sequence, and distinguish normal stage differences from unusual spatial patterns. Small fields may contain only a limited number of observations that are not mixed with boundaries, tracks or other crops. Quality screening should account for this limitation rather than treating every nominal pixel as an independent view of onions. Comparable acquisition and processing conditions strengthen the interpretation of changes between reporting periods.
Field visits should record crop stage, stand continuity and the condition of the soil and bed surfaces contributing to the image. Sample across the mapped range, including areas that look typical, and preserve the location and date of each observation. A general reflectance pattern cannot directly establish bulb production or a named cause of stress. Those claims need appropriate measurements, while a transparent inspection-priority workflow can be evaluated through the outcomes of practical scouting visits.
Start with the intended claim
Validation tests how well a product represents a defined quantity under specified conditions. It differs from calibration, which adjusts an instrument or model using reference information. Begin with the claim the product needs to support: class membership, surface reflectance, canopy height, water quality, or an area estimate. State the required spatial scale, time interval, and acceptable error. These choices determine the reference method and sample design. Collecting convenient field points first and deciding later what they prove usually leaves important conditions untested.
Reference measurements also contain uncertainty. Positioning error can place a sample in the wrong satellite pixel, a point instrument may not represent a heterogeneous footprint, and measurements made on different dates can describe real surface change. Record instrument calibration, sampling protocols, timestamps, coordinates, observation conditions, and uncertainty where available. For interpreted labels, retain the evidence used, class rules, and ambiguous cases. Agreement among interpreters is useful to examine, but agreement alone does not guarantee that the shared interpretation is correct.
Keep the assessment independent
Separate validation data from observations used to fit thresholds, train models, select features, or tune processing. If assessment sites repeatedly influence revisions, they have become development data and a fresh independent assessment is needed. Spatially clustered samples can make performance look stronger when neighboring training and test locations share conditions. Use a design that reflects the target population and report its limitations, including inaccessible areas or environmental conditions missing from the sample.
Choose metrics that match the product. For categorical maps, examine a confusion matrix and errors for individual classes, not just overall agreement. For continuous measurements, report bias and a suitable error measure alongside sample count and the range of reference values. Plot residuals against relevant conditions to expose failures hidden by an average. For area estimates, carry the probability sampling design through estimation and uncertainty calculations. Preserve rejected matchups and their reasons so that quality screening remains auditable. A useful validation report explains where the product works, where evidence is sparse, and which errors matter for the intended decision. Publishing reference protocols and assessment code makes later comparisons more credible and helps others reproduce the evaluation.
Design an evaluation that reflects deployment
A model evaluation should resemble the conditions in which the model will be used. Start by identifying which relationships the development data share: nearby locations, repeated observations of the same unit, common collection teams, or measurements from a single period. These relationships can make records less independent than they appear. An evaluation that ignores them may test the ability to recognise familiar conditions rather than the ability to operate in a new setting.
Choose a split strategy that matches the intended transfer. Holding out locations, seasons, operational units, or acquisition conditions can answer different questions. Document which question each split is designed to address, and keep the held-out evidence separate while selecting features and tuning the method. Repeatedly inspecting the same evaluation set to improve a model gradually makes that set part of development, even if it retains its original label.
Report results at a scale that is meaningful to the user. A strong average can hide poor performance for important subgroups or unusual conditions. Examine error patterns and retain examples of both successful and unsuccessful predictions. Where the evidence covers only a limited range of conditions, describe that range explicitly. The evaluation then becomes a practical guide to suitable use, further testing, and the collection of additional reference observations.
Deliver quality flags that users can interpret
Quality information should describe the condition of an observation in terms that remain useful through the workflow. A single valid or invalid label may be sufficient for some calculations, but operational users often need to know why an observation was qualified. Distinguish acquisition limitations, processing failures, uncertain reference matches, and conditions that make interpretation ambiguous. These categories point towards different responses and should not be merged without a clear purpose.
Define a small, documented set of flags with stable meanings. Record whether a flag applies to a complete scene, a geographic unit, an individual measurement, or a derived result. A unit containing a mixture of qualified and unqualified observations may also require a coverage measure. Include the rule used to combine flags when summarising data so that a favourable average does not conceal a substantial fraction of unusable input.
Test the flags with the people receiving the product. Ask what they would do differently for each state and whether the terminology supports that choice. Keep the original flags available even if the interface presents a simpler summary. When flag definitions change, version the rules and review their effects on historical comparisons. Useful quality communication makes the condition of the evidence visible at the moment when a user decides how much confidence to place in the result.
Bring the evidence back to the decision
Expansion should be guided by measured workload, visible quality conditions, and an understanding of which tasks require expert attention. Test the workflow under representative data volumes before treating a successful demonstration as evidence of operational capacity. Preserve a route for investigating individual observations even when most processing is automated. A scalable service keeps its results traceable and useful as it grows, with clear choices about where to automate, where to review, and where to invest in additional evidence.
Sources & further reading
Explore the bhumisat spacecraft concept and our vision towards 2030, or return to the journal.



