Attribute combination: RGB blending and classification

The RGB blend recipe

• Pick three attribute volumes. Spectral 15 / 30 / 60 Hz is the most famous combination, but any three attributes that you want to compare side-by-side will do.
• Normalize each to a 0–255 range. Typical practice: clip each volume to its own [0, 3·RMS] (or some percentile range) and linearly map to [0, 255]. Each channel gets its own normalization so that meaningful variation in any one channel is visible regardless of absolute magnitude.
• Combine into one image. At each pixel: red = R-channel value, green = G-channel value, blue = B-channel value. The eye reads the resulting hue as the relative balance of the three attributes.

That is it. The technique is mathematically trivial — the magic is entirely in the choice of inputs.

Reading colours: the spectral RGB blend

The classic combination is low / mid / high frequency. For F3-style data: 15 Hz to red, 30 Hz to green, 60 Hz to blue. Pixel colours then encode the spectral signature of the rock at that location:

• Pure red — Strong low frequency, weak mid and high. Indicates thick beds (which tune at low frequencies) or attenuated zones. Sometimes a gas-charged or heavily-absorbing layer.
• Pure green — Strong mid frequency only. Typical for layers whose dominant tuning matches the survey's peak frequency — the "ordinary" reflectors that fall in the centre of the spectrum.
• Pure blue — Strong high frequency. Indicates thin beds (which tune at high frequencies) or sharp interfaces. Often delineates fine stratigraphic features that mid-frequency mapping misses.
• Yellow (red + green) — Energy at low and mid, weak high. A "fat" spectrum without sharp interfaces — thicker bedded sequences.
• Cyan (green + blue) — Mid + high but no low. Often marks zones with no big thick reflectors but lots of thin-bed detail.
• Magenta (red + blue) — Low + high but no mid. Less common but interesting — sometimes a tuning gap effect or a localized acquisition / processing artefact.
• White (all three) — Balanced spectrum across all frequencies. Indicates rock with broadband reflectivity — typically clean sand-shale sequences with no strong tuning effects.
• Black — No energy at any of the three. Either silent zones (mute, above seabed) or zones outside the data's usable bandwidth.

The remarkable thing is that channels and other stratigraphic features often appear in distinct colours on a spectral RGB blend even when they're hard to see on any single iso-frequency image — because their characteristic thickness picks one specific frequency and shifts the balance toward that channel's colour.

Exercise — The F3 RGB story

• The widget starts with the classic spectral blend on F3 (R = 15 Hz, G = 30 Hz, B = 60 Hz). Slide the time slider down through the volume. You should see colour patterns that change with depth, telling you the spectral signature is not uniform.
• Around 1100–1400 ms, look for distinct colour patches — yellow zones, cyan zones, mottled regions. Each colour represents a different spectral signature, hence (probably) a different lithology or thickness regime.
• Change the Red channel dropdown to RMS amplitude. Now you have an "amplitude / mid-freq / high-freq" blend. The visual transforms: amplitude is now red rather than spectral content, so bright amplitude zones glow red regardless of their spectrum. This is a useful blend for spotting bright spots whose frequency content is also distinctive.
• Now change the Green channel to Coherence and Blue to Mean curvature. You've built the structural blend: amplitude (red), coherence (green = HIGH coherence, since the value goes up where rocks are similar), curvature (blue, with absolute value of the signed curvature). Most of the image will be greenish (high coherence is the norm); fault zones will stand out as dark or magenta lineaments (low coherence + nonzero curvature); folds will glow blue.
• Switch the View dropdown to Inline. RGB blends work as vertical sections too, though they're less iconic on inlines than on time slices. The colour pattern across depth tells you how the spectrum changes vertically — a useful sanity check on whether the volume's frequency content is reasonably stationary or has obvious depth-dependent loss (the deeper you go, the redder the rock should become if the data has typical absorption-driven spectral evolution).

Useful blend recipes (besides spectral)

The R / G / B assignment is just convention. Pick combinations that answer the question you have:

• Amplitude / Frequency / Coherence — a "general purpose" blend. Bright reflectors with characteristic frequency on continuous rock light up cleanly; faults break the pattern.
• Coherence / Dip / Curvature — the structural blend. Faults appear where coherence drops; folds where curvature increases; tilted regions where dip is high. Combined: you see the entire structural fabric of the rock at one depth.
• Three different envelope volumes from different time gates — useful for time-lapse comparisons (4D seismic) where you want to see what changed between vintages.
• Inversion-derived attributes (Acoustic Impedance, V_p/V_s, density) — the rock-physics blend. Each colour combination corresponds to a specific lithology / fluid combination, calibrated against well control. This is the standard quantitative-interpretation deliverable.

The constraint isn't the technique — it's the human eye. We can integrate three colour channels reliably; four or more we cannot. So pick your three carefully.

Beyond RGB: crossplotting and classification

RGB blending is the simplest way to combine multiple attributes. Two more advanced approaches scale further:

• Crossplotting. For each voxel, plot attribute A vs attribute B as a point on a 2D scatter (or three attributes as a 3D scatter). Identify clusters in the crossplot space. Then "highlight" each cluster's voxels back in 3D — the crossplot becomes a tool for selecting all rock that shares a particular attribute combination. Standard for AVO interpretation: cross-plot intercept vs gradient, identify the AVO Class III cluster, then map where in the volume those voxels live.
• Classification volumes. Apply k-means, Gaussian Mixture Models, or self-organizing maps to the multi-attribute vector at each voxel. The output is a discrete cluster ID per voxel — a classification volume. Each cluster represents a distinct multi-attribute "facies" that may correspond to a geological unit, a lithology, a fluid state, or simply a noise category. Modern machine-learning workflows extend this to convolutional / transformer architectures for facies prediction directly from the seismic.

These techniques are out of scope for this section's widget but rest on exactly the same foundation: every attribute volume we've built in §6.1–§6.4 is an input to multi-attribute interpretation. The more attribute families you have computed, the richer your classification can be.

Common pitfalls of multi-attribute display

• Colour blending obscures absolute values. A pixel rendered "green" only tells you "G channel is dominant"; you don't know whether G is at 30% of its clip range or 95%. For quantitative interpretation, return to the single-attribute view to read absolute magnitudes.
• Each channel's normalization is independent. Different blends will look completely different even on the same data, because each channel's clip range is computed from its own statistics. The colour at one pixel is meaningful only relative to the choice of all three normalizations.
• Three is the eye's limit. You cannot meaningfully blend four attributes into one image. Common workarounds: blend three, then make a second image with a different three; or use crossplots and classification as above.
• Colour-blind users. Red–Green colour-blindness is the most common form. Spectral RGB blends with R+G discrimination at their core may be inaccessible. Consider a complementary palette (e.g., yellow / cyan / magenta) or provide a single-attribute fallback view.
• The blend is only as good as its inputs. If one of the three attributes is low-quality (noisy, biased, computed with the wrong window), the blend inherits that quality. Always validate each input attribute on its own before combining.