C5 · Publication Volume 15

Spatial, Spectral, Radiometric and Temporal Resolution

resolution trade-offs and common interpretation errors

Learning goals

This lesson treats resolution as a four-dimensional design problem. The learner should be able to distinguish pixel spacing, footprint and resolving power; separate spectral sampling from bandwidth; relate bit depth to radiometric sensitivity without confusing them; describe revisit, acquisition opportunity and usable temporal density; and select a product whose combined resolutions match the geological decision.

No dataset has a single “resolution.” Improving one dimension often changes swath, signal, storage, revisit or another dimension. A useful specification states what difference must be detected, at what scale, under what noise and how often.

Spatial sampling, footprint and resolving power

Pixel spacing is the distance between output grid centres. Ground sampling distance describes detector sampling projected onto the surface. The instantaneous field of view and system point-spread function describe how energy from neighbouring ground locations contributes to a sample. Orthorectification and resampling create an output grid but cannot restore detail absent from the measurement.

A feature smaller than a pixel can influence a value if its contrast is strong, but it is not thereby resolved. Resolving two adjacent features requires adequate modulation transfer, sampling, contrast and signal-to-noise ratio. A nominal 10 m grid derived from a 30 m observation remains limited by the original support. Conversely, a coarse pixel may still detect a large uniform spectral change.

Edges, narrow veins, roads, shadows and small disturbed areas produce mixed pixels. Their apparent width depends on footprint, resampling and orientation relative to the grid. State minimum mapping unit separately from pixel size.

Spectral sampling and diagnostic resolution

Spectral sampling interval is the spacing between channel centres. Spectral resolution is related to the width and shape of channel response. Contiguous narrow channels can characterise absorption position and shape; a few broad bands can separate large spectral contrasts but may average narrow features. Oversampling a broad instrument response does not create finer spectral resolution.

A diagnostic feature must be considered after convolution with response functions, atmosphere and noise. Its centre may fall between bands, its shoulders may share one band, or a water-vapour interval may remove information. Feature separability depends on more than channel count: wavelength calibration, signal-to-noise ratio, stray light and correlated noise are material.

Compare sensors in measurement space. Resample reference spectra to each response, add realistic noise and test whether competing materials remain separable. Do not count nominal bands as independent variables without inspecting covariance.

Radiometric and temporal resolution

Radiometric resolution describes the instrument and product's ability to represent small energy differences. Stored bit depth sets the number of possible codes, but effective resolution also depends on full-well capacity, gain, noise, calibration, quantisation, saturation and scaling. A 16-bit file does not guarantee 16 bits of useful information. Report signal-to-noise ratio at relevant radiance levels and wavelengths.

Temporal resolution is often described by revisit, but usable observation density can be lower because of cloud, illumination, acquisition strategy, view limits, quality screening and product latency. Repeat observations from different sensors can increase cadence only after geometry, spectral response, calibration and processing are harmonised.

Change may be impulsive, gradual, seasonal or episodic. Sampling must resolve the process and its baseline. Two dates cannot separate a one-off event from ordinary seasonal variability unless external evidence constrains the history.

Resolution trade-offs and task design

Smaller spatial footprints collect less energy unless aperture, dwell time or illumination changes; narrower bands also reduce photons. Fine spatial and spectral sampling can therefore reduce signal-to-noise ratio, swath or revisit. Compression and onboard processing add further trade-offs. There is no universally superior product.

Define the decision unit first. A regional lithological pattern may favour broad coverage and consistent radiometry. A narrow alteration zone may need fine spatial support and diagnostic short-wave infrared bands. A monitoring task may favour frequent, harmonised observations over the finest single scene. A terrain-lineament task may depend on elevation accuracy and point density rather than optical bands.

Build a table with required feature width, spectral feature, minimum contrast, maximum cloud gap, dynamic range and acceptable false-decision rate. Score candidate products against these requirements and identify which requirement is not met.

Aliasing, mixing and uncertainty

Spatial aliasing occurs when sampling is too coarse for scene variation, causing high-frequency patterns to appear as misleading lower-frequency patterns. Spectral aliasing can occur when sparse bands miss or misplace feature shape. Temporal aliasing can turn seasonal cycles into apparent trends. Radiometric clipping replaces a range of true values with one saturated code.

Mixing is physical and mathematical. A pixel may contain several materials, illumination states and subpixel slopes; a band integrates wavelengths; a temporal composite integrates dates. Linear spectral mixing can be a useful approximation for area mixtures under common illumination, but intimate mixtures, multiple scattering and nonlinear atmosphere–surface interactions violate it.

Quantify sensitivity by degrading a higher-resolution synthetic truth with response kernels, noise and resampling. Compare detection, boundary position and class confusion. The degraded product, not the original truth, is what the interpretation must support.

Worked synthetic example

A synthetic 18 m wide alteration strip crosses a 10 m grid obliquely. The optical point-spread function has substantial energy beyond one grid cell. The strip affects three to five pixels, but only the central pixels approach its pure spectrum. After resampling to a 30 m grid, no pixel is pure; the feature remains detectable in a diagnostic ratio but its mapped width becomes 60 m.

Weekly acquisitions are nominally available, yet cloud and shadow leave one usable observation per month. A two-month seasonal vegetation change has the same sign as the alteration-sensitive ratio. The spatial dataset can flag a corridor, but the time series cannot attribute the ratio without vegetation control and field validation. Reporting a 10 m boundary would overstate both spatial and causal resolution.

Interpretation and decision.

Report the scale at which the evidence supports a claim. Use “a mixed-pixel corridor is detected” rather than “an 18 m body is mapped” when footprint and registration do not support the boundary. For spectra, state which feature attributes survive sensor convolution. For time series, state observation dates actually used and gaps after masking.

Resolution is decision-specific. Detectability may be sufficient to prioritise a broad check even when geometry or identity is unresolved. A higher-stakes material claim needs independent data at matching support. The next observation should target the limiting resolution dimension rather than simply add more of the same product.

Practice and audit checklist

Create a synthetic scene with narrow, broad and low-contrast targets. Convolve it with two spatial kernels, two spectral response sets and two noise levels, then sample it on two time grids with missing observations. Measure detection, boundary error and confusion. Explain which loss is irreversible.

Audit questions: Is pixel size confused with footprint? Is bandwidth available? Does stored bit depth exceed effective dynamic range? Are saturated values masked? Is revisit distinguished from usable cadence? Does the minimum mapping unit exceed one pixel? Were mixed pixels and resampling kernels recorded?

Four resolution dimensions jointly control which geological differences remain observable
Four resolution dimensions jointly control which geological differences remain observable

Sources