C5 · Publication Volume 15

Multispectral Image Interpretation

bands, composites, indices and visual interpretation

Learning goals

This lesson turns a small set of broad spectral bands into testable geological observations. The learner should be able to select bands and composites from a physical hypothesis; distinguish display stretch from measurement; calculate ratios and normalised differences with masks and uncertainty; use tone, texture, shape, pattern and context without overclaiming; and design validation that tests alternative causes of an image anomaly.

Multispectral interpretation works because materials, cover and illumination produce different broad spectral and spatial patterns. It is powerful for reconnaissance and context, but broad bands rarely identify one mineral uniquely. The task is to extract repeatable evidence at the scale and spectral support of the sensor.

Bands, composites and spectral hypotheses

A single band is a response-weighted measurement over a wavelength interval. A three-channel composite maps selected band values to display red, green and blue. A natural-colour assignment approximates human visible perception; a false-colour assignment deliberately moves non-visible information into visible channels. Neither is intrinsically more truthful. The scientific question determines the useful assignment.

Begin with predicted relative behaviour. If hypothesis A expects high near-infrared vegetation response and hypothesis B expects exposed iron-bearing regolith with stronger red than blue reflectance, choose bands that make those differences measurable. Write expected signs before viewing the target. A composite selected after finding a visually pleasing anomaly invites confirmation bias.

Inspect every constituent band separately and plot representative spectra before interpreting composite colour. The same colour can arise from many combinations of channel values, and two display channels may be strongly correlated. Record band identifiers, wavelength response, processing level, scale and masks with the composite.

Display stretch, colour and visual comparability

A display stretch maps numeric values to screen intensity. Linear minimum–maximum, percentile, standard-deviation and histogram-based stretches alter contrast but not source measurements. Separate stretches per scene can make unchanged surfaces look different; a global stretch can hide subtle variation in one scene. Clipping bright or dark tails may erase valid extremes or conceal saturation.

Colour tables impose perception. Sequential palettes suit ordered magnitude; diverging palettes suit signed departure from a meaningful centre; categorical palettes suit discrete classes. Rainbow palettes can create artificial boundaries and unequal perceptual emphasis. Legends must state the physical value before display transform.

For comparison, use fixed limits or a documented reference transform and show saturation counts. Preserve an unstretched quantitative view or provide access to the underlying values. Image sharpening, pan fusion and local contrast enhancement change spatial support and must be labelled as derived display products, not original spectral measurements.

Ratios, normalised differences and feature contrast

A ratio R=x/y can suppress common multiplicative illumination and highlight relative spectral slope, but it becomes unstable when y is small and does not remove additive path radiance or shadow. A normalised difference N=(x-y)/(x+y) is bounded when positive inputs are valid, yet remains sensitive to offsets, noise, saturation and mixtures. Always apply scale and offset before calculation.

For independent uncertainties, a first-order ratio variance is approximately


\sigma_R^2\approx \frac{\sigma_x^2}{y^2}+\frac{x^2\sigma_y^2}{y^4},

with a covariance term required when errors are correlated. Resampling, common atmospheric correction and shared calibration commonly create correlation. Mask low denominators, saturation, cloud, shadow and invalid aerosol retrievals before computing ratios.

An index is a model, not a material detector. Its numerator and denominator encode an expected contrast. Test the index on alternative surfaces, season, moisture, terrain and sensor response. Report its valid domain and avoid transferring thresholds across products without recalibration.

Tone, texture, shape, pattern and context

Visual interpretation uses tone or colour, texture, shape, size, shadow, pattern, association and site. Geological units may express through drainage, landform, fracture traces, soil development or vegetation as well as spectrum. Linear features may be structures, roads, field boundaries, acquisition seams or shadows. Circular patterns may be intrusions, impact features, drainage arrangements or processing artefacts.

Texture depends on scale and display. Quantitative texture measures require a declared window, direction, quantisation and boundary treatment. A rough appearance at one zoom is not an intrinsic rock property. Morphological continuity can support a structural hypothesis only if it persists across illumination, sensor and terrain representations.

Context should constrain, not predetermine, interpretation. Compare the image feature with mapped topography, regolith domain, drainage, known disturbance and independent geophysics or field observations. Keep a table of features predicted by each hypothesis and mark observations that conflict.

Quantitative extraction and validation

Define regions or transects before extracting values where possible. Avoid selecting only the brightest centre of an anomaly and a convenient dark background. Include representative within-class variation, edge mixtures and negative controls. Account for spatial autocorrelation when splitting training and validation data; neighbouring pixels are not independent replicates.

Report distributions, not only means. Plot band values, ratios and spectra with valid-pixel counts, masks and acquisition conditions. Test robustness to stretch, mask dilation, atmospheric quality, terrain illumination and spatial aggregation. A feature that disappears under a one-pixel registration perturbation is weak evidence for a narrow boundary.

Validation must target the claimed property. Field photographs alone may confirm exposure but not mineral identity. Spectral measurements need compatible geometry and calibration; material claims may need petrography, diffraction or chemistry. Use independent sites and record where the product correctly returns “unknown.”

Worked synthetic example

A synthetic scene contains three exposed units, patchy dry vegetation and a bright road. Unit A has a gently rising visible-to-near-infrared continuum. Unit B has stronger red response and lower near-infrared response. Unit C resembles B in broad bands but has a narrow absorption that the sensor does not resolve. A red–near-infrared ratio separates A from B and C over bare pixels, while the road produces the strongest ratio because of its different overall spectrum.

The ratio therefore maps a broad spectral contrast, not Unit B. Shape and road association reject the infrastructure response; vegetation and shadow masks remove other confounders. B and C remain indistinguishable. A defensible output is a corridor compatible with two exposed-material hypotheses and a request for imaging spectroscopy or field mineralogy, not a mineral-class map.

Interpretation and decision. Describe what is observed before naming a cause: band values, ratio sign, spatial continuity and mask support. Then list compatible geological and non-geological explanations. The decision may be to prioritise field checking, to exclude disturbed pixels or to seek a sensor with a diagnostic band. State what result would falsify the preferred explanation.

Practice and audit checklist

Using a synthetic five-band scene, define two competing material hypotheses, predict their band ordering, build two composites and one physically motivated index, and evaluate three negative controls. Compare fixed and scene-specific stretches without changing underlying calculations. Produce a table separating measurement, visual cue, inference and validation need.

Audit questions: Were scale and offset applied? Are invalid denominators masked? Are display and analysis products separate? Was the composite chosen from a prior hypothesis? Are shape and texture evaluated at a declared scale? Does validation match the claimed property? Are thresholds transferable, or merely fitted to this scene?

Bands, composites, indices and contextual tests convert broad spectral measurements into conditional geological evidence
Bands, composites, indices and contextual tests convert broad spectral measurements into conditional geological evidence

Sources