Define what the label claims
Classification assigns observations to categories. In satellite analysis, those categories might describe vegetation, exposed ground, water or built surfaces. The apparent simplicity hides a crucial distinction: land cover describes physical surface characteristics, while land use describes human purpose. A grass-covered surface could be pasture, parkland or an airfield margin. The image may support a cover label without supporting a confident statement about ownership, management or use.
Begin with a written class definition and rules for difficult cases. Does “tree cover” include orchards? Does temporary floodwater count as water? How much of a pixel must be occupied by a feature before it receives the label? CORINE Land Cover and ESA WorldCover offer established classification contexts, but their legends, spatial support and intended applications differ. Borrowing a class name does not automatically make a new analysis comparable with either product.
Labels carry field assumptions
A supervised model learns from labelled examples, so training data are part of the measurement chain. Labels derived from field visits, aerial interpretation or an existing map each carry their own limits. A reference polygon drawn several years earlier may no longer describe the scene. A broad polygon may contain hedgerows, tracks and ponds that do not fit its dominant label. Document the source date, interpretation rule and positional uncertainty instead of treating every training example as equally reliable.
Represent the conditions in which the model will be used. An urban model trained in central London may struggle with settlement patterns around Nairobi. Vegetation structures, roofing materials, soils and seasonal conditions can differ substantially. This is a domain-shift problem, not necessarily a software fault. Diverse training examples help, but they do not replace independent tests in the target geography. Unknown or ambiguous cases should remain visible rather than being forced into a convenient category.
Choose a model that can be examined
Random forests, gradient-boosted trees and neural networks can all support remote-sensing classification. Convolutional models and other spatial architectures may exploit texture and neighbourhood context; simpler models may be easier to diagnose with limited labels. There is no universal winner. A fair comparison uses the same evaluation framework, realistic withheld data and an account of computational demands. Complexity is justified by improved task performance and useful behaviour, not by the novelty of the model name.
Input features matter just as much. Spectral bands, seasonal summaries, terrain variables and radar measurements may help separate classes, but each brings preprocessing assumptions. The Normalized Difference Vegetation Index, or NDVI, summarises a contrast between near-infrared and red reflectance. It can help describe vegetation response, but it is not a species label, a direct biodiversity measure or proof of ecological health. Keep the interpretation of each feature narrower than the environmental story it might suggest.
Do not let nearby pixels mark the test
Neighbouring image samples are often similar. Randomly splitting pixels from the same field between training and test sets can make evaluation appear stronger than performance in a new area. Spatially separated evaluation, geographic holdouts and time-separated tests address different deployment questions. Decide which question matters: filling gaps within a familiar landscape, transferring to another region or predicting a future season. Report the separation strategy alongside the scores so readers understand what has actually been tested.
Overall accuracy can conceal weak results for rare classes. A model that maps the dominant background well may still miss small wetlands or confuse sparse settlements with bare soil. Inspect the confusion matrix, per-class precision and recall, and the spatial distribution of errors. If reference sampling intentionally overrepresents uncommon classes, population-level estimates may require sampling weights. A score without its sampling design and reference definitions is not enough to support a policy conclusion.
Publish uncertainty with the map
A model score is not automatically a calibrated probability of correctness. High confidence can occur on unfamiliar inputs, systematic label errors or observations outside the training distribution. Review confidence calibration when probabilities are used in decisions, and consider an abstention rule for cases requiring human assessment. Boundary zones and mixed pixels deserve particular attention because a single discrete label can conceal multiple surface types within the observation footprint.
Finally, state what the classified map is for. A regional screening layer may be useful for directing field visits while remaining unsuitable for parcel-level enforcement. The Open Geospatial Consortium provides standards relevant to exchanging geospatial information, but a technically interoperable file still needs clear semantics and validation. Publish the legend, input period, model version, evaluation limits and known failure modes together. A defensible category is one whose meaning remains understandable after it leaves the analyst’s workstation.