Relative Assessment of Greenness with RGB Indices: How to Compare Leaf Colour Between Plants, Treatments and Dates

What relative greenness assessment means

There are two ways to use a number that comes out of a leaf photograph, and they demand completely different levels of rigour.

Absolute assessment asks: what is the chlorophyll content of this leaf? Answering it honestly requires a calibrated instrument, a species-specific and condition-specific calibration curve, and a chain of evidence back to laboratory values. An RGB index alone cannot do this, and any workflow that implies it can is overselling.

Relative assessment asks: is this leaf greener than that one, and by how much on a consistent scale? This is a comparison question, and RGB indexes are very good at it — provided the two leaves were photographed with the same instrument, under comparable light, and scored with the same formula.

An RGB greenness index is a ruler with no zero and no unit. It measures differences beautifully and levels not at all.

What an RGB index actually extracts from a photo

Understanding why relative comparison works — and absolute reading does not — starts with what happens between the shutter and the number.

1. Segmentation. Computer vision separates leaf pixels from background, shadow, tray, hand and soil. Everything downstream depends on this being clean.
2. Pixel filtering. Pixels lost to specular highlight, deep shadow, blur or boundary mixing are excluded. The share that survives is the valid-pixel ratio.
3. Per-pixel index computation. Each surviving pixel's R, G and B values are pushed through the index formula, producing a per-pixel index map — the false-colour images throughout this article.
4. Robust summarisation. The per-pixel map is collapsed to a statistic per leaf. Petiole Pro reports the pixel median, which is far less sensitive to a handful of extreme pixels than a mean.
5. Aggregation. Leaf-level values are aggregated to plant, plot or treatment level — and this is the first point at which a comparison exists at all.

Two of the indexes Petiole Pro computes matter for this article, because they are the pair examined in the experiment below.

DGCI (Dark Green Colour Index) is derived from the HSB/HSV colour space and is the index most widely used in turfgrass science and nitrogen research. It rewards a pixel for being the right hue of green, vividly rather than washed out, and dark rather than pale. Because hue is separated from brightness, a change in light intensity perturbs mostly one component instead of smearing across three channels. Our field guide to measuring greenness with DGCI in Petiole Pro covers the calibration-plate workflow behind it.

AGRI (Advanced Green Relative Index) is a customised RGB vegetation index used within the Petiole Pro pipeline. It is built directly from the visible bands and tuned for close-range leaf imagery rather than for canopy or aerial scale. Note what its name already concedes: it is a relative index. Its job is to place leaves on a common scale within a batch, not to report an absolute pigment concentration.

Why the absolute number does not travel

Five independent factors change the index value of a leaf that has not changed at all. Each of them is neutralised by a relative design and none of them is neutralised by quoting a bare number.

1. The camera is a stylist, not an instrument

Smartphone cameras apply automatic white balance, auto-exposure, tone curves, saturation boosts and increasingly aggressive computational processing designed to make photographs look pleasant. Two phone models pointed at the same leaf disagree; so does one phone in two modes. A calibration reference in the frame anchors colour, but it anchors it to that capture chain.

2. Illumination is a variable you are also measuring

Direct midday sun, overcast sky, open shade and greenhouse LEDs produce four different RGB triplets from one leaf. Colour temperature shifts the balance between channels; intensity shifts brightness; a passing cloud shifts both inside a single session. Without a relative reference, a "greenness trend" can be a weather trend.

3. The leaf is three-dimensional

Curvature produces brightness gradients that have nothing to do with chlorophyll. Waxy cuticles throw specular highlights — blown-out patches where colour information is simply gone. Leaf angle relative to the light source changes the recorded value of an unchanged pigment load.

4. Every index has its own arbitrary scale

DGCI, AGRI and TGI are normalised differently, span different numeric ranges, and respond differently to the same perturbation. There is no conversion factor. A leaf can sit high in one index's distribution and mid-pack in another's.

5. Biology varies without stress

Leaf age, node position, sun versus shade leaves, and cultivar genetics all shift greenness independently of the treatment you are testing. A relative design controls these by holding them constant across the compared groups; an absolute reading has no way to separate them from the effect.

None of this makes RGB indexes weak. It makes them conditional. Fix the conditions, compare within them, and the measurement is repeatable to a degree that visual scoring never achieves.

Evidence: 37 leaves, two indexes, r = 0.316

The argument above is not theoretical. Petiole R&D Lab experiment 008, run on 26 August 2026, provides a clean demonstration using a dataset collected by a Petiole mobile app user and processed automatically on the Petiole Pro Web platform.

The distributions

Already the practical lesson is visible. If you had reported "mean DGCI 0.41" and someone else had reported "mean AGRI 0.38", the two numbers would look comparable and would be describing the same leaves through incompatible lenses. The scales overlap numerically by coincidence, not by design.

The correlation

DGCI and AGRI correlated at r = 0.316 across the 37 leaves: a positive but weak relationship, with roughly 10% of the variance in one index explained by the other (R² ≈ 0.10). Both formulas ran on identical pixels from identical photographs. Nine-tenths of what each index saw, the other did not. That is not an error — it is two formulas responding to different characteristics of the same image, exactly as designed.

The consequences for study design follow directly:

• An index is part of the method, not a detail of reporting. "Greenness increased by 8%" is meaningless without naming the index. Declare it before you collect data and do not change it.
• Rankings are index-specific. A leaf near the top of the DGCI distribution can sit mid-pack in AGRI. If your conclusion is "line 14 was the greenest", it is a conclusion about line 14 under that index.
• Agreement is information. Where two indexes rank a leaf similarly, confidence is higher. Where they diverge sharply, something in that image — a highlight, a shadow, an odd exposure — usually deserves a human look.
• Never average indexes together. Combining scales with different ranges and different sensitivities produces a number that tracks neither.

Leaf-to-leaf spread is the signal, not the noise

Across the 37 leaves, greenness varied considerably from leaf to leaf. It is tempting to read that as measurement noise to be suppressed. It is mostly biology: leaves differ in age, node position, light history, pigmentation and condition. A single leaf is a poor estimate of a plant, and a single plant is a poor estimate of a plot.

Relative assessment turns this from a problem into a design requirement. Three consequences:

• Sample enough leaves that you can see a distribution. The comparison you are making is between two distributions, not two points. With spread this wide, a handful of leaves per treatment cannot separate a real effect from sampling luck.
• Standardise leaf position ruthlessly. Youngest fully expanded leaf, or a fixed node — chosen once and held. Greenness varies systematically along a stem, so comparing different positions compares position, not treatment.
• Report spread alongside centre. A treatment median with an interquartile range communicates what a bare mean hides, and it is the honest form of a result built on variable material.

Four ways to make greenness relative

"Relative" is not a single technique. Four designs cover almost every practical question, and they differ in what they control for and what they demand of your capture protocol.

1. Control ratio — the strongest design

Leave an over-fertilised reference strip in the field, photograph it in the same session as the plots under test, and express every plot as a fraction of that reference. This is the logic that made the nitrogen sufficiency index a standard tool with chlorophyll meters, and it transfers directly to RGB indexes. Because the reference and the sample share camera, light, day and operator, almost every confound cancels in the ratio. If the light was unusually flat, both numbers moved; the ratio did not.

2. Within-batch z-score — for screening

When there is no natural control — a breeding nursery, a germplasm screen, a survey — standardise against the batch itself. Each leaf becomes "how many standard deviations from this trial's mean". Z-scores make different trials visually comparable and are the natural input to a selection threshold. They do assume the batch mean is a meaningful anchor, so they are unsuited to a batch where most entries share one treatment effect.

3. Percentile rank — for robustness

Given how skewed index distributions can be, a rank-based summary is often more honest than a parametric one. Percentile rank discards the arbitrary scale entirely and keeps only the ordering, which is the part of the measurement that actually transfers. It is the right choice when the distribution has a long tail, when outliers are suspected image artefacts, or when the decision is genuinely "select the top 10%".

4. Time delta — for dynamics

Comparing the same plants to themselves is the cleanest form of relative assessment, because genetics, node position and leaf age are held constant by construction. It is also the most demanding on protocol: the same device, the same hour of day, the same weather category, the same leaf position, the same calibration reference. A stay-green phenotyping study, a senescence curve or a post-application response check all live here. For canopy-scale versions of the same idea, see our work on digital grapevine canopy analysis.

These designs compose. A stay-green trial might use a time delta on plot medians, each plot expressed as a ratio to an irrigated control, and selection made on percentile rank within date. What none of them permit is comparing a raw index value to one from another study.

The quality gate: valid-pixel ratio

Every greenness value carries an implicit question: how much of this leaf did you actually measure? Specular highlights, deep shadow, motion blur and boundary pixels all produce pixels that must be excluded. The valid-pixel ratio is the fraction of pixels inside a detected leaf region that survived filtering and contributed to the index.

A leaf scored on 40% of its pixels is not the same kind of observation as one scored on 75%, and averaging them together as equals quietly corrupts a treatment mean. In a relative design this matters twice over, because a systematic difference in capture quality between two groups — one photographed in brighter sun, say, with more blown highlights — produces a difference in index values that has nothing to do with the plants.

Set the threshold before you look at the results. A valid-pixel ratio floor of 0.60 is a reasonable starting point for close-range leaf imagery, and experiment 008 suggests most well-captured leaves clear it comfortably. Apply the same floor to every group, record how many leaves each group lost to it, and report that count. If one treatment lost twice as many leaves as another, you have found a capture problem before it became a conclusion.

A step-by-step protocol for a defensible comparison

The following sequence turns the principles above into a repeatable field procedure. It assumes Petiole Pro with a Greenness Calibrating Plate, but the logic applies to any RGB pipeline.

1. Write down the comparison before you photograph anything. Name the groups, the reference, the index, the leaf position and the sample size. A greenness dataset collected without a declared comparison almost never becomes one afterwards.
2. Choose one index and fix it. DGCI is the default choice for nitrogen and turf-adjacent work because of its established relationship with chlorophyll-meter readings. Record the choice in your method.
3. Establish a reference. An over-fertilised strip, an untreated control, or a baseline date. Every later number will be expressed against it.
4. Standardise capture. One device for the whole study. Diffuse light or your own shadow, never direct sun. A consistent hour of the day. A clean lens and an unstained calibration plate. Full plate — ArUco markers and colour references — inside every frame.
5. Standardise the leaf. Youngest fully expanded leaf, or a fixed node, on every plant in every group.
6. Sample for a distribution. Enough leaves per group that the spread is visible, given how much leaf-to-leaf variation experiment 008 showed. Replicate across plants and plots, not just across leaves of one plant.
7. Capture the whole comparison in one session. Everything that will be compared directly should share a session. If a study spans days, re-photograph the reference each day and compare within day.
8. Process in batch. Run the full image set through one pipeline with one parameter set, so no group receives different treatment.
9. Apply the quality gate. Filter on valid-pixel ratio at the threshold you declared in step 1. Record the exclusions per group.
10. Compute the relative quantity. Control ratio, z-score, percentile rank or time delta — chosen in step 1, not selected after seeing which one gives the nicer result.
11. Report the method with the number. Index, device, lighting, leaf position, sample size, quality threshold, exclusions, and the relative form used. This is what makes the result reproducible; it is also what makes it citable.

Worked example: two treatments, one afternoon

Consider a biostimulant trial. Two blocks, treated and untreated, plus an over-fertilised reference strip. The question is whether the treated block held greenness better than the untreated one three weeks after application.

The wrong way. Photograph 5 leaves from each block on whatever afternoon suits, note that treated averaged DGCI 0.44 and untreated 0.41, and conclude a 7% improvement. Nothing here is defensible: the sample is too small for the observed leaf-to-leaf spread, the two blocks may have been shot in different light, no quality gate was applied, and the difference is expressed on a scale that has no external meaning.

The relative way. In one afternoon session on one phone, in open shade, photograph 30 youngest fully expanded leaves per block plus 30 from the reference strip, calibration plate in every frame. Batch-process all 90 images with one parameter set. Drop leaves below 0.60 valid-pixel ratio and record how many went from each group. Express each block as a ratio of its median DGCI to the reference median. Report:

Illustrative figures, shown to demonstrate the reporting form rather than to report a specific trial. The second version answers the question. The ratio column is portable in a way the DGCI column is not: it can be compared to next season's ratio, to another site's ratio, and to a colleague's ratio, because the reference absorbs the camera, the light and the index scale. The exclusion counts show that the quality gate hit all three groups similarly, which is itself part of the evidence.

Seven mistakes that break a relative comparison

Running it in Petiole Pro

The workflow splits naturally between capture and analysis.

In the field, the Petiole Pro mobile app measures greenness non-destructively from an ordinary smartphone photograph with the Greenness Calibrating Plate in frame. Because the measurement does not damage the leaf, the same plants can be revisited for a time-delta design — which is the strongest way to run a senescence or stay-green study.

At the desk, batch processing is what makes a distribution-sized sample practical. Point the pipeline at a folder, and every image is processed with one parameter set: leaves detected and segmented, indexes computed per pixel, per-leaf medians and valid-pixel ratios recorded, and per-object data exported for filtering and statistics in R, Python or a spreadsheet. Experiment 008's 37 leaves went through exactly this route on the Petiole Pro Web platform. A full account of the capabilities is on the Petiole Pro features page, and the measurement workflows it supports are set out in the Petiole Pro use cases.

Greenness is rarely the only trait worth extracting from the same photograph. The same segmentation that feeds a greenness index also supports necrotic leaf area assessment for damage quantification, and the capture discipline described here is the same discipline set out in our guide to photographing leaves for accurate measurement in the field. If you are setting up from scratch, the Petiole Pro store has the app and the calibration plate.

What relative greenness can and cannot tell you

Trustworthy measurement means being explicit about limits. A relative design makes RGB greenness indexes genuinely useful; it does not make them a spectrometer.

Relative greenness assessment can:

• Rank leaves, plants, plots, treatments and cultivars quickly and repeatably within a session.
• Detect change over time in the same plants, non-destructively, at a cadence manual scoring cannot sustain.
• Express a plot as a fraction of a well-fed reference, which is portable between seasons and sites in a way a raw index value never is.
• Flag suspect measurements, through the valid-pixel ratio and through disagreement between indexes.
• Reveal within-leaf spatial patterns — interveinal chlorosis, edge burn, tip dieback — through per-pixel maps.
• Deliver enough replicates for real statistics, because batch throughput is high.

Relative greenness assessment cannot:

• Report chlorophyll or nitrogen in absolute physiological units without local calibration against laboratory or meter values for your crop and conditions.
• Make values comparable across studies, cameras, lighting regimes or indexes. This is the central claim of this article and it has no workaround.
• Explain why a leaf is pale. Nitrogen deficit, sulphur deficit, iron chlorosis, root disease, waterlogging and natural senescence all reduce greenness.
• Recover information a photograph lost. A blown-out highlight contains no colour to measure.
• Rescue a comparison whose groups were photographed under different conditions. No post-hoc correction substitutes for one session, one device, one protocol.
• Substitute for area, count or damage measurements. Greenness answers "how green", not "how much leaf" or "how much damage".

Read with those boundaries in mind, a relative greenness design does exactly what plant health monitoring needs: it turns a subjective impression into a repeatable comparison, fast enough and cheap enough to collect at the scale real decisions require.