Begin with the feature, not the file
Imagine checking whether a wetland edge has moved. Before downloading imagery, define what counts as an edge, how far it might move and during which season the comparison matters. A narrow drainage ditch, a flooded meadow and an entire catchment require different observation strategies. AI cannot recover information that the sensor did not record. A detailed-looking output can still be an elaborate estimate built from observations too coarse for the intended claim.
Write the question as a measurable relationship: surface water extent during comparable seasonal conditions, for example, rather than simply “wetland health.” At the Somerset Levels, water management and temporary inundation complicate any single scene. In the Okavango Delta, the timing of incoming floodwater matters as much as local rainfall. These examples illustrate why geographical context belongs in the acquisition plan rather than in an explanation added after classification.
Read resolution in several dimensions
Spatial resolution describes the ground sampling of an image, but a pixel is not a neatly isolated parcel. Surface contributions mix, and neighbouring features influence the signal. Sentinel-2 provides selected bands at 10, 20 and 60 metres; resampling a 20-metre band to a 10-metre grid does not create new 10-metre detail. Landsat 8 and Landsat 9 offer a long-running observation framework with 30-metre multispectral bands. Their usefulness depends on the scale and history of your question.
Spectral resolution concerns the wavelengths measured. Near-infrared and shortwave-infrared observations can help distinguish vegetation, moisture and exposed ground, but no band is a universal detector of environmental condition. Temporal resolution concerns opportunities to observe, not guaranteed usable scenes. Radiometric behaviour and processing consistency matter too. A model trained on one product may respond differently when values, scaling or atmospheric correction change, even if the landscape has remained stable.
Cloud is only one kind of obstruction
Optical observations are affected by cloud, haze and shadows. Snow, smoke and low solar angles create additional problems. A scene with little cloud overall can still have cloud directly over the study area. Inspect the actual area of interest and quality masks rather than relying only on scene-wide metadata. In mountainous terrain such as the Alps, topographic shadow and slope orientation can produce patterns that resemble differences in vegetation or land cover.
Sentinel-1 radar provides a different observation route because it does not depend on sunlight and can observe through cloud. That advantage does not make radar a direct substitute for an optical photograph. Backscatter responds to surface roughness, moisture, geometry and other properties. Flooded vegetation may behave differently from open water, while buildings can create strong returns. Select acquisition geometry and processing carefully before asking a model to combine radar with optical inputs.
Match time to the process
A crop field changes throughout its growing cycle without changing its land use. A river changes width with discharge. A coastal flat changes with the tide. Across the Mekong Delta, a date chosen without regard to cropping and inundation can make a normal seasonal state look unusual. For a comparison, record the observation dates and the environmental conditions that make them meaningfully comparable. A cloud-free composite is useful, but its pixels may represent different dates.
Compositing therefore trades one problem for another. It can reduce cloud contamination while obscuring short-lived events or mixing phenological stages. State the compositing window, selection rule and missing-data treatment. If the task concerns a particular flood, a seasonal median may erase the event you need to study. If it concerns persistent land cover, that same median may suppress distracting short-term variation. There is no acquisition recipe independent of the analytical purpose.
Leave an acquisition record
Keep the source collection, product identifiers, processing level, acquisition times, coordinate reference system, quality filters and resampling choices. Record excluded scenes and the reasons for exclusion. The European Space Agency, the Copernicus programme and the United States Geological Survey publish product documentation that should accompany any workflow. Access through a convenient platform does not remove the need to understand the underlying instrument and product.
Finish with an observation sufficiency statement: what can be distinguished, what remains mixed and what needs another source. If a field boundary is narrower than the usable observations, say so. If clouds leave a critical period unseen, preserve that gap instead of silently filling it with certainty. A defensible acquisition plan is not the one with the most images. It is the one that makes the relationship between evidence, scale and uncertainty explicit.
Next dispatch
When pixels become categories