Image acquisition is the first step of any remote-sensing workflow, and the quality of every later output, from thematic maps to machine-learning models, depends on it. At this stage the sensor, platform, time window and processing level are selected from the practical question the project has to answer rather than from the newest or finest data on offer. The aim is a final image set that is geometrically well georeferenced, radiometrically comparable with the other scenes in the same study, and usable in terms of cloud cover and seasonal conditions. What is handed over at the end of this step is not raw data but a set of corrected scenes with complete metadata and an explicit note of each scene’s limitations.
Choosing the Sensor and Platform
Passive optical sensors, such as the multispectral instruments on the Landsat and Sentinel platforms, record energy reflected from the surface in visible, near-infrared and shortwave-infrared bands and support work on vegetation, water, soil and built-up areas. Active synthetic aperture radar, such as the C-band instrument carried by Sentinel-1, supplies its own illumination, so it is independent of sunlight and largely independent of cloud, which makes it the better choice for ground instability, surface moisture and heavily clouded regions.
The platform matters just as much: satellite imagery for wide areas with regular revisit, UAVs where very fine detail is needed over a small site, and multi-decade archives for long-term trend work. The choice among these follows first from the spatial extent of the study area and from the size of the smallest feature that has to be separated.
- Wide coverage with regular revisit and a long archive: multispectral satellite imagery.
- Penetration of cloud, or acquisition at night and in poor illumination: active radar sensors.
- Centimetre-scale detail over a small site: UAV photogrammetry.
- Elevation and imagery in one product: stereo pairs or lidar data.
Resolution and Its Trade-offs
Four kinds of resolution act at the same time in acquisition, and improving one is normally paid for by another. Spatial resolution sets the size of the smallest feature that can be separated, spectral resolution the number and width of the bands, temporal resolution the interval between two successive acquisitions, and radiometric resolution the number of distinguishable brightness levels.
Finer spatial resolution in optical sensors usually means narrower bands and therefore weaker spectral resolution, which makes more sensitive indices, such as the soil-adjusted family, harder to apply reliably. Shorter revisit gives more cloud-free opportunities inside a cropping window or around an event, but it also raises storage volume and processing load. Radiometric resolution, in turn, decides how much of a subtle brightness difference survives into a quantitative model.
- Fine spatial resolution: small features are visible, but spectral detail drops and data volume rises.
- High spectral resolution: better separation of surface types, but coarser pixels.
- Short revisit interval: fast events can be tracked, at the cost of more storage and processing.
- High radiometric resolution: finer brightness steps and better quantitative modelling.
Archive Search and New Tasking
There are two routes to the imagery: searching what already exists in an archive, or tasking a new acquisition. In an archive search the area of interest is defined as a polygon, together with a date range, a cloud-cover ceiling, an acceptable off-nadir angle, a minimum sun elevation and a list of acceptable instruments.
In tasking, the acquisition window is agreed with the provider, and two practical constraints apply: requests usually close well before the pass, and weather can waste a booked pass entirely. Tasking therefore makes sense for time-critical events or unusual geometric requirements, while long-term monitoring is normally served better from the archive.
The order also fixes whether the product is delivered at one processing level or another, which radiometric range and date window are acceptable, and, where relevant, how the acquisition geometry such as a stereo pair or a specific incidence angle should be configured. Recording these decisions in the project file gives a reference point when data sufficiency is questioned later.
Cloud, Shadow and the Seasonal Window
Cloud cover is the leading reason a scene is discarded in humid and mountainous regions, and cloud shadow also invalidates otherwise healthy pixels, so a separate mask should be delivered rather than left to the user. For multi-date studies, drawing all scenes from one season and, as far as possible, one phenological stage prevents a difference in vegetation condition from being misread as a change.
Snow, dust, smoke and high reflectance over water surfaces also affect radiometric quality and introduce systematic error into time series. Locking the analysis window around planting, harvest or the period before and after an event is part of the study design and should be settled before the order is placed.
- Deliver a cloud and cloud-shadow mask with each scene and document how it was produced.
- Set the cloud-cover threshold from the needs of the analysis, not simply as low as possible.
- Keep season and growth stage consistent across the dates being compared.
- Assess snow and bright soils separately for their effect on class separability.
Processing Levels and Corrections
Raw sensor data passes through a chain of processing before delivery: radiometric correction to convert digital numbers to radiance or top-of-atmosphere reflectance, geometric correction to bring the scene onto a reference grid, and, at a higher level, atmospheric correction to reach surface reflectance.
For quantitative comparison between dates and for spectral indices, surface reflectance is the correct input; top-of-atmosphere data is not suitable because of atmospheric scattering and because solar geometry differs between dates. Georeferencing uses the rational polynomial coefficients supplied with the scene and, where the application demands it, ground control points and shared tie points.
- Radiometric correction and consistent scaling across all scenes in the set.
- Atmospheric correction for surface reflectance in multi-date analysis.
- Geometric correction with residual shift checked against control points in the study area.
- Masking of invalid pixels: cloud, cloud shadow, snow and saturation.
- Uniform data range and clipping to the study boundary.
Delivery Formats and Metadata
Imagery is normally delivered as georeferenced GeoTIFF, with a Cloud Optimized GeoTIFF copy for archiving and publication, since tiled and progressive reading makes it far more efficient in a web service. Some compressed licensed products arrive as JPEG2000, which requires a suitable codec library to open.
Metadata should record the scene footprint, acquisition date and time, sun and sensor angles, scale and offset coefficients, no-data value and the geometric model coefficients. For organisational projects the description follows spatial metadata standards and is delivered alongside a history of the processing steps, so the dataset remains traceable years later.
- GeoTIFF for analysis, Cloud Optimized GeoTIFF for publishing and archiving, JPEG2000 for compressed deliveries.
- Rational polynomial coefficients or a geometric model report with every scene.
- Coordinate reference system and EPSG code stated in metadata and in file naming.
- A step-by-step processing log so results can be reproduced.
Summary
The quality of a remote-sensing project depends more on choosing and preparing imagery correctly at this stage than on the algorithms used at the end. Establishing the study boundary, the smallest feature of interest, the event window and the required processing level before ordering prevents expensive rework.
The output of this stage is a set of corrected scenes, valid masks, complete metadata and a written note of limitations, and it is the direct input to the image analysis stage.