Remote Sensing Labs
Browse all docs

Digital Imagery

A remotely sensed image is not a photograph in the everyday sense. It is a structured grid of measurements: each cell records how much energy a sensor detected from a small patch of the Earth’s surface. Reading imagery well means understanding what that grid actually contains, how the stored numbers relate to real physical quantities, and why every value carries some uncertainty.

Pixels, grids, and metadata

A digital image is a regular grid of pixels, and each pixel of a gridded product represents a fixed area on the ground — a 10-meter or 30-meter square, for example — holding one number for every band the sensor measured. The grid has a set number of rows and columns, but the array of numbers only becomes a map once it is tied to the Earth by georeferencing metadata: a coordinate reference system plus an affine transform that fixes where the corner pixel sits and how large each pixel is. Without that metadata the pixels have values but no location. The pixel size sets the grid’s spacing on the ground; the finest detail the image can actually distinguish also depends on the sensor’s optics and on any resampling, which the Resolution concept develops further. Metadata additionally records the acquisition time, the sensor and its bands, and the processing history, all of which you need before you can compare one image to another.

From digital numbers to physical values

The raw value stored in a pixel is often a digital number: an integer the sensor’s electronics produced when they quantized the incoming energy. A digital number is not itself a physical quantity — it is a coded measurement whose meaning depends on the sensor’s calibration. To interpret imagery consistently, digital numbers are converted to calibrated physical values such as at-sensor radiance, top-of-atmosphere reflectance, or, after atmospheric correction, surface reflectance. Calibrated values are expressed in physical units rather than sensor-specific codes, so they can be compared across scenes, dates, and sometimes across different sensors. This is why raw-looking imagery and calibrated reflectance can resemble each other yet must not be used interchangeably: only calibrated values support meaningful comparison, indices, and change analysis. Knowing which product you are holding — raw digital numbers, radiance, or reflectance — is a prerequisite for every later step.

A fourth state is easy to miss: many reflectance products are stored as scaled integers, so the stored number must be multiplied by a scale factor and shifted by an offset recorded in the metadata before it means reflectance. The factors differ between products and have changed between processing versions of the same product, so they are read from the metadata, never assumed.

Scene footprints, nodata, and valid observations

Each acquisition covers a finite footprint on the ground, bounded by the sensor’s swath and the satellite’s position at the time. When that footprint is stored in a rectangular grid, the corners and edges include cells that fall outside the actual observation, and those cells are marked as nodata rather than given a real measurement. Masks for clouds, shadow, saturation, or sensor gaps are often applied the same way, turning those pixels into nodata for analysis, although many products record them in separate quality flags instead. Treating nodata as if it were a genuine measurement — reading it as a value of zero, for instance — is a common and damaging mistake, because it silently biases averages, indices, and classifications. A dependable analysis therefore begins by separating valid observations from nodata, so that every later calculation runs only on pixels that actually recorded the surface.

Image collections and repeated observations

A single image is one snapshot in time. Because satellites revisit the same ground again and again, imagery naturally accumulates into collections: many scenes of the same area captured on different dates, and often by more than one sensor. Thinking in collections rather than isolated images is much of what makes remote sensing powerful, since it lets you build cloud-free composites, follow seasonal cycles, and detect change over time. It also imposes a discipline: the scenes in a collection must share, or be resampled onto, a common grid, calibration, and coordinate system before their pixels can be compared cell by cell. The Time Series concept builds directly on this way of seeing imagery.

Where distortion, noise, and uncertainty come from

No measurement is perfect, and imagery carries several kinds of error. Geometric distortion comes from terrain relief, the sensor’s viewing angle, and small errors in the platform’s position and attitude, so a pixel can sit away from its true ground position until the image is orthorectified — corrected with an elevation model — and then aligned on a common grid, a Coordinate Systems concern. Radiometric error has several sources: random noise from the sensor’s own electronics, imperfect calibration between detectors and over time, the changing geometry of sun and view, and atmospheric effects, as the atmosphere scatters and absorbs energy on its way to the sensor. Either way a recorded value is an estimate rather than an exact truth. Clouds, haze, shadow, and mixed pixels — cells that straddle more than one surface type — add further ambiguity. None of this makes imagery unusable; it means each pixel value should be read as a measurement with uncertainty attached. Recognizing where that uncertainty comes from is what separates a defensible result from an accidental one.

Taken together, these ideas are the foundation for everything else. Once you can read an image as a georeferenced, calibrated, partly masked, and uncertain grid of measurements — and as one member of a larger collection — the concepts of Spectral Bands, resolution, and coordinate systems all build on top of it.

Sources