GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-05
A classifier can only answer the question represented by its training and evaluation. That boundary becomes particularly important when the output concerns damaged buildings. A confident label may describe one of several studied damage patterns while saying nothing about whether a building is safe to enter or occupy.
Research first submitted on 24 February 2023 by Vivien Zahs and colleagues investigates a random-forest model trained with virtual laser scanning data and applied to real photogrammetric point clouds. The evaluation distinguishes three grades: heavy damage, extreme damage and destruction, using pre-event and post-event observations. These categories define a specific research task. They do not constitute a complete building-safety assessment. Original research
Read the missing categories
For an architectural research review, begin with a category table rather than the headline accuracy. List each tested label and the evidence used to assign it. Alongside that table, list situations outside the demonstrated scope, including forms of deterioration or hidden damage that the observations may not reveal.
GAD proposes an explicit unresolved category in any experimental reporting workflow. An unfamiliar case, inadequate capture or inconsistent input should be allowed to remain unclassified. Forcing every building into the closest available class can make a narrow model appear more comprehensive than it is.
This is a proposal for evaluation design, not a claim that the published classifier implements such a category. A qualified structural professional and the relevant authorities must make safety, access and occupancy decisions through the appropriate assessment process.
Separate change from capture difference
Before interpreting a changed point-cloud region, examine the comparability of the observations. Capture coverage, registration and occlusion can influence apparent differences. Preserve the original inputs and record the preprocessing steps so a reviewer can identify whether an unexpected result began with the observations or the classification.
Use difficult cases when examining transfer from synthetic training to real data. Include geometry unlike the simulated examples and regions with incomplete observations. Report the failure pattern and available evidence, rather than explaining away a poor result through a more favourable overall score.
The errors also have different consequences. Missing a concerning case and unnecessarily flagging a case both matter, but they do not create equivalent decisions or workloads. A research evaluation should present those outcomes separately and explain who reviews each unresolved case.
Keep the output in its proper role
An experimental map can help organise evidence for specialist review. Its legend should state the studied categories, input dates and important limitations. Avoid colours or labels that imply unrestricted access or a certified safety status.
GAD has not run this model or independently verified its reported performance. The practical lesson is about the boundary of an AI claim: a method trained to distinguish particular severe damage patterns should remain accountable to that task. Uncertainty must stay visible precisely where a reader might otherwise mistake a classification for permission to act.