A classification accuracy figure is incomplete until the proposal identifies which records were evaluated. Performance on training examples does not answer the same question as evaluation on held-out examples. For a UG25 mapping inquiry, request separate data-role labels, a sample register and an intended-use statement before comparing headline scores. No reported percentage alone establishes that a proposed aircraft and processing package suits your project.

Read the caption before the percentage

A land-cover proposal presents a polished map and an accuracy result in large type. The result may be correctly calculated, yet still answer a narrower question than the buyer assumes. Was the score calculated on the records used to build the classifier? On a separate evaluation set? On a different site? A missing caption can make these distinct evidence types look interchangeable during a purchasing discussion.

The first useful request is not a higher number. Ask the supplier to identify the evaluated population and explain how those records entered the workflow. This is a procurement question about the evidence being offered, not an accusation that the model is defective. A modest, clearly scoped result can be more decision-useful than an impressive figure whose origin cannot be inspected.

An official example keeps the roles separate

Google's Earth Engine documentation for Classifier.confusionMatrix states that this method computes a confusion matrix from training data. Its example separately classifies validation samples and computes a validation error matrix. The distinction is sufficient for a buyer's first question: which evaluation does the proposal actually report?

The example does not establish a suitable sampling design for every drone-mapping project. It also does not demonstrate that UG25 includes, integrates with or has been evaluated using that software. Treat the documentation as evidence for the terminology, then ask the proposed processing specialist to justify the actual evaluation design. A familiar API name is not a substitute for a project-specific evidence record.

Give every sample a declared role

A practical sample register should make it possible to trace a record from its source to its assigned role. Request identifiers, the reference label, the source of that label and whether the record was used for training, tuning or final evaluation. The exact format can be simple. What matters is that the recipient can distinguish a known role from an assumption made after seeing the result.

Ask how revisions are handled. If a label is corrected, or a sample is excluded, the report should identify the reason and the affected evaluation. Do not allow an unexplained change in the sample population to disappear behind a refreshed chart. The buyer needs to understand whether two versions evaluate the same question, not merely whether the newest file has a later date.

Define the intended population

The purchasing team should state where and for what decisions it expects to use the classified map. An evaluation for one named area is not automatically evidence for every future site. Ask the specialist to describe the population to which the result is intended to apply, the excluded conditions and the evidence still needed for a wider claim. Keep the answer attached to the result rather than buried in an unrelated appendix.

This article does not prescribe a universal split ratio or a minimum sample count. Those choices need a defensible design for the actual question. Our editorial recommendation is narrower: require the designer to make the choices explicit and assign a reviewer who can evaluate them. A buyer should not select a convenient ratio from an online example and treat it as a complete validation plan.

Inspect the records behind a disputed result

Imagine a fictional handover in which a buyer cannot reproduce a table because the delivered sample register lacks the role field. The supplier has supplied all the coordinates, but the recipient cannot tell which records belong in the reported comparison. The appropriate response is to clarify the register and evaluation instructions, not to guess a partition that happens to produce a similar number.

This is original buyer analysis, not a reported customer case. Its practical lesson is to test a small receiving workflow before final acceptance. Have the intended reviewer identify a record's source, label, role and inclusion in the relevant summary. Record unresolved items as exceptions. That exercise checks the transparency of the handover; it should not be described as independent scientific validation of the whole map.

Separate evidence review from model selection

A purchasing comparison should not encourage bidders to optimize only the most visible figure. Ask each bidder to supply the same evidence categories and declare missing information. If one offers only a training result while another offers a documented evaluation on reserved records, do not place the two percentages in an unlabeled best-to-worst column. First establish whether they are comparable claims.

Keep commercial terms separate as well. An aircraft, a processing service, a classification deliverable and an evaluation review can be different parts of a proposal. Ask who owns each responsibility and which outputs are included. A lower aircraft price cannot explain an omitted validation service, just as a more detailed report does not establish that an aircraft configuration has been approved for the customer's intended operation.

Do not let the application label replace an output definition

A map described as agricultural, environmental or infrastructure-related still needs a precise target definition. Ask which categories the buyer actually needs to distinguish and how uncertain or excluded cases will appear. The report should help the recipient identify what can be used, what needs review and what was not assessed. A broad application name cannot resolve a mismatch between the buyer's categories and the supplier's labels.

Derived terrain products raise a related, but different, question about the meaning of output values. The UG21 flow-accumulation guide separates counted contributing cells from a water-discharge claim. Both purchasing situations reward precise labels, but their evidence must remain separate. A classification accuracy result does not validate a hydrological interpretation simply because both appear in the same mapping report.

Train the receiving team to ask the right question

The employee reading the proposal may not have designed a classifier. A short, ground-only learning exercise can help them identify missing evidence without asking them to act as the technical approver. Provide a worked example of a clearly labeled report, then ask them to inspect a changed example and explain what is missing. The UVH-RTF learning-package discussion examines that distinction between seeing a solution and producing independent reasoning.

Do not turn that exercise into a claim of professional competence. Its purpose is to improve the intake conversation and direct technical questions to the appropriate reviewer. Keep a route for saying the evidence is not understood or not sufficient. A procurement process is stronger when an unresolved data-role question remains visible instead of being converted into a confident ranking.

What to confirm about UG25

The approved UG25 product page identifies the model for a configuration inquiry. It does not establish that a classifier, processing subscription, sample collection service or evaluation package is included. Request a proposed configuration and scope of supply that addresses the actual mapping objective. For specifications and integration facts not confirmed in that scope, contact us for configuration details.

US and UK teams can use these evidence questions without assuming their projects have identical operational or professional requirements. The article is about report review, not flight planning or regulatory approval. Establish those requirements separately and make sure the supplier's response identifies which questions concern the aircraft, which concern the sensor and which concern the data service.

Common purchasing questions

Is a training score useless?

No. It describes a stated evaluation on training records. The problem is presenting it as if it answers a different question. Keep its role clear and request the additional evidence needed for the intended purchase decision.

Does held-out automatically mean representative?

No automatic project-wide conclusion follows from that label. Ask how the records were selected, what population they represent and what limitations remain. The responsible specialist should explain the design rather than asking the buyer to infer it from a filename.

What should be compared first?

Compare definitions, data roles, reference evidence and scope before ranking results. Record non-comparable items as such. A transparent gap is more useful than a synthetic comparison that gives different evidence types the same heading.

Request an evidence-shaped proposal

Prepare your intended map categories, receiving workflow and decision scope, then ask for separate training and evaluation summaries with the supporting sample register. Review the VTOL and fixed-wing collection as a starting point for aircraft discussion, not as evidence that a classification outcome is guaranteed.

Send the brief through the UNITED UAV inquiry page and request written confirmation of the proposed UG25 configuration, service inclusions and unresolved evaluation requirements. The most useful next answer is not an unexplained accuracy percentage. It is a clear account of what was evaluated and whether that evidence addresses your intended use.