Digital Greenness Indices for Tomatoes: TGI, DGCI, NGRDI and NDPI Compared on a Fertiliser Trial Row

What is a digital greenness index?

A digital greenness index is a formula that converts the red, green and blue values of a pixel in an ordinary camera image into one number that rises as the plant gets greener. Because leaf greenness is driven mainly by chlorophyll, and chlorophyll by nitrogen supply, the index works as a fast, non-destructive proxy for plant nitrogen status. Computed for every plant pixel in an image, it produces a greenness map and a set of summary statistics: the median, the mean and the percentiles.

Greenness indices are the visible-light cousins of NDVI. NDVI needs a near-infrared band and therefore a multispectral camera; RGB indices work with the camera already in your pocket. The trade-off is that a phone camera is not a calibrated instrument. It applies its own white balance, exposure and colour processing, so RGB index values are reliable for comparing plants photographed together and much less reliable as absolute numbers. We covered that limitation in depth in our guide to relative greenness assessment with RGB indices.

There are dozens of published RGB indices. They differ in three ways that matter in practice:

• Colour space. Some work directly on R, G and B (TGI, NGRDI). Others first convert to hue, saturation and brightness (DGCI), which separates "which colour" from "how bright".
• Normalisation. Normalised-difference indices such as NGRDI are bounded between −1 and +1 and partly cancel overall brightness. Unnormalised indices such as TGI keep a wider, open-ended scale.
• What they were built for. TGI was designed to track leaf chlorophyll at canopy scale; DGCI was designed to rate turfgrass colour; NGRDI was developed for vegetation monitoring from red and green bands.

Why measure greenness in a tomato fertiliser trial?

Nitrogen is the nutrient a tomato grower watches most closely. Too little, and the plant runs short of chlorophyll: leaves turn paler, starting with the older leaves lower on the plant, and growth slows. Too much, and the plant puts on lush, dark foliage at the expense of flowers and fruit. Fertiliser trials exist to find the programme that sits between the two, and leaf colour is one of the first places a difference between programmes shows up.

The usual ways of checking are slow or limited:

• Visual scoring is quick but subjective. The eye adapts to light and compares colours poorly across a row or between days.
• A chlorophyll meter (SPAD) is objective, but it reads a few square millimetres of one leaf at a time, so covering a canopy takes dozens of clicks per plant.
• Leaf tissue analysis gives actual nitrogen content, but it is destructive, takes days and costs money per sample.

An RGB greenness index sits in between. One top-down photo or video frame covers hundreds of thousands of leaf pixels across several plants, and the index is computed in seconds. The question this article answers is a practical one: which of the common indices is actually useful for comparing tomato plants that received different fertilisers, and what has to be controlled for the comparison to be fair?

Which four indices did we compare?

All four were computed from the same segmented plant pixels in every frame, so any difference between them comes from the formula, not from the data.

TGI: Triangular Greenness Index

TGI was introduced by Hunt and colleagues (2013) to estimate leaf chlorophyll from visible bands only. It measures the area of a triangle drawn between the red, green and blue reflectances: the deeper the chlorophyll absorption in red and blue relative to the green peak, the larger the triangle. The published form is TGI = −0.5 × [190 × (R − G) − 120 × (R − B)], using the band centres of 670, 550 and 480 nm. TGI is not bounded, so its numeric range depends on how the image values are scaled.

DGCI: Dark Green Colour Index

DGCI was introduced by Karcher and Richardson (2003) to quantify turfgrass colour from digital images. It converts each pixel to hue, saturation and brightness and combines them as DGCI = [(Hue − 60) / 60 + (1 − Saturation) + (1 − Brightness)] / 3. For green pixels it falls between 0 and 1, with higher values meaning a darker, deeper green. Our field guide to DGCI covers it in detail.

NGRDI: Normalised Green–Red Difference Index

NGRDI compares the green and red channels as a normalised difference: NGRDI = (G − R) / (G + R). It is commonly traced back to Tucker's (1979) work on red and green band combinations for vegetation monitoring. Values run from −1 to +1. Positive values mean green dominates red; values near zero or below indicate yellowing, brown tissue, soil or stems. We used it to separate healthy from necrotic tissue in our guide to necrotic leaf area.

NDPI: a normalised-difference RGB index

NDPI, as computed in our pipeline, is a second normalised-difference index built from the visible bands. Like NGRDI it is bounded between −1 and +1 and partly cancels overall brightness, but it weights the bands differently, so it responds to a different part of the colour signal. In this dataset its frame medians sat around 0.41, well above NGRDI's 0.14.

How the four compare

How was the data collected?

The set-up was deliberately simple, close to what a grower or agronomist would do on a routine crop walk.

• Plants. A single row of tomato plants of the same variety, grown in bags on drip irrigation. Plants along the row received different fertilisers. Because the variety was the same, differences in leaf colour are attributable to growing conditions rather than genetics.
• Capture. A smartphone held above the canopy, pointing straight down, recorded a continuous video while walking the length of the row.
• Frames. 139 frames were extracted from the video, each 1080 × 1920 pixels (portrait).
• Segmentation. In every frame, each pixel was classified as plant or background (paving, grow bags, drip line). Plant pixels made up between 27.3% and 65.3% of a frame, 43.5% on average: about 0.9 million plant pixels per frame and 125.3 million in total.
• Indices. TGI, DGCI, NGRDI and NDPI were computed for every plant pixel, giving a heat map per index per frame.
• Statistics. For each frame and index we recorded the pixel count, minimum, maximum, mean, median and the 5th, 10th, 25th, 75th, 90th and 95th percentiles.

What do the four index maps look like on the same plants?

Even before any numbers, the maps show that each formula sees a different plant:

• TGI is low (blue) across most of the shaded inner canopy and high (red) on individual sunlit leaflets and leaf edges. It is the most sensitive to how light falls on each leaf.
• DGCI is the most uniform. Most of the canopy sits in a narrow yellow-green band, with orange patches where the foliage is darkest. Blue spots mark pale or glare-affected pixels.
• NGRDI is highest in the dense centre of each plant and lower on the thin, upper shoots at the edges, which is a first hint that it tracks canopy density as well as colour.
• NDPI produces a fine-grained speckle with little large-scale structure, picking up texture at the scale of individual leaflets.

For a fertiliser trial this is the first lesson: a "greener plant" is a statement about an index. The same leaflet can be in the top tenth of the frame under TGI and unremarkable under DGCI.

Results at a glance

The table summarises each index across the 139 frames. "Frame median" is the median of all plant pixels in a frame; the columns then describe how those 139 frame medians varied.

CV = standard deviation of the 139 frame medians divided by their mean. The first and last thirds are frames 0–45 and 93–138. "After correcting for coverage" removes the part of each index explained by a straight-line relationship with canopy coverage, and expresses the remaining change as a percentage of the index mean.

Bottom line: TGI showed the largest and most persistent change along the row; DGCI was the most stable and the only index independent of canopy coverage; NGRDI looked responsive until coverage was accounted for, which halved its apparent change.

A word of caution on the CV column. CV depends on where an index's scale starts, so a 13% CV on TGI and a 1.4% CV on DGCI are not directly a measure of which index is "more sensitive to nitrogen". They show how much each index moved across the same plants, which is what decides whether a fertiliser difference will stand out from the noise.

How did each index change along the row?

Reading the panels from top to bottom:

• Canopy coverage peaked at 65% around frame 15 and fell towards the end of the row, averaging 46.4% in the first third, 44.0% in the middle and 40.1% in the last third.
• TGI drifted down along the row, from a first-third average of 3.45 to 3.08 in the middle third and 2.74 in the last. Its lowest stretch, frames 87–104, had medians of 2.2–2.5, against 3.5–4.2 around frames 15–25.
• DGCI stayed within a band of 0.04 for the entire row. The first, middle and last thirds averaged 0.525, 0.530 and 0.527.
• NGRDI rose and fell with canopy coverage, peaking around frame 54 (0.176) and reaching its lowest values in the last 15 frames (0.083–0.117), where coverage was also lowest.
• NDPI followed a similar but much flatter pattern to NGRDI, averaging 0.417, 0.417 and 0.399 across the three thirds.

The heat maps make the disagreement between the two indices concrete. At frame 21, TGI is lifted by hundreds of sunlit leaflets and leaf edges scoring in the red and yellow; its median is 4.16. By frame 96 those bright points have largely gone, the canopy reads blue, and the median falls to 2.21. Over the same frames DGCI moves the other way, from 0.513 to 0.542. The plants did not become both paler and darker green: each index is weighting a different property of the same pixels.

Because fertiliser differed between plants along this row, a gradient along it is consistent with a treatment effect. It is equally consistent with a change in light along the walk, or with plants at the end of the row being smaller. Without treatment labels the two cannot be separated, which is why the protocol below starts with labelling plants before filming.

Note also how smooth the lines are. The correlation between each frame and the next was 0.92 for TGI, 0.86 for DGCI, 0.93 for NGRDI and 0.87 for NDPI. Consecutive frames from a walking video overlap heavily, so 139 frames are far fewer than 139 independent observations.

Does canopy coverage bias a greenness index?

Yes, for three of the four indices in this dataset. Canopy coverage, the share of the frame classified as plant, explained 23% of the frame-to-frame variation in TGI, 29% in NGRDI and 25% in NDPI, and essentially none (r = +0.02) in DGCI. Frames with more foliage in view scored greener on the RGB-difference indices, independently of any change in leaf colour.

Why would the amount of foliage change the colour reading? Several mechanisms push in the same direction:

1. Mixed edge pixels. A sparse canopy has more leaf edges per square centimetre. Edge pixels blend leaf with paving or grow bag, which pulls the green–red balance towards neutral.
2. Leaf layering. Where foliage is dense, light passes through several leaf layers before reaching the camera's view of the inner canopy, and the green signal is stronger.
3. Plant part mix. Thin, upper shoots and stems make up a larger share of a sparse frame. They are paler and redder than mature leaves.
4. Plant size itself. Coverage is partly a measure of plant vigour, and vigour can be a genuine fertiliser effect. Correcting for it removes some real signal along with the bias.

The last point is why coverage should be recorded and reported rather than silently "corrected away". In a fertiliser trial, the honest analysis shows both: the raw index difference between treatments, and the difference after accounting for canopy size. If an effect survives the correction, as TGI's did here, it is more likely to be a colour difference and not just a size difference.

Do the four indices agree with each other?

Partly. The table shows correlations between the 139 frame medians of each pair of indices (Pearson r; Spearman's rank correlation was within 0.07 of every value).

Three patterns stand out:

• TGI, NGRDI and NDPI form a family. They are all built from differences between RGB bands, and they correlated at r = 0.48–0.59. Moderate, not strong: even within one family, an index explains at most about a third of the frame-to-frame variation in another.
• DGCI runs slightly against TGI and NDPI. Frames that TGI rated greener, DGCI tended to rate marginally less green (r = −0.27). Working in hue, saturation and brightness, DGCI rewards dark, saturated green; TGI rewards the depth of red and blue absorption, which is strongest on bright, sunlit leaves.
• NGRDI bridges the two. It correlated positively with every other index, including DGCI (r = +0.57), which makes it a useful cross-check between the two families.

The practical rule follows directly. Never average indices together, and never switch index halfway through a trial. If two indices from different families both rank treatment A above treatment B, that is strong evidence. If only one does, look at the images before drawing a conclusion.

This mirrors what we found on single leaves in an earlier experiment, where DGCI and AGRI correlated at only r = 0.316 across 37 leaves (see relative greenness assessment with RGB indices). Whether you measure one leaf or a whole canopy, the choice of index is part of the method.

Why is the spread inside one frame so large?

Within a single frame, greenness varies far more than it does between frames. For every index, the typical 25th–75th percentile range of plant pixels in one frame was 6 to 15 times larger than the standard deviation of the frame medians along the whole row. Leaf-to-leaf and pixel-to-pixel variation dominates; differences between parts of the row are small by comparison.

That spread is not error. A tomato canopy genuinely contains:

• young, pale leaflets at the growing tips and older, darker leaves below;
• sunlit upper surfaces next to leaves in their neighbours' shade;
• leaf veins, petioles and stems, which are paler than the leaf blade;
• specular glare on leaves angled towards the sun.

It has two consequences for a fertiliser trial. First, a single photo of a single leaf says very little about a plant; summarise many pixels from many leaves. Second, the percentiles carry information the median hides. Nitrogen deficiency in tomato tends to show first on older leaves, so a treatment effect may appear in the lower tail (the 10th or 25th percentile) before it moves the median.

Out-of-range and undefined pixels: mean or median?

Every index produced some values that look wrong at first sight:

• DGCI pixel values ranged from −0.11 to 2.20, outside its nominal 0–1 scale. The formula assumes green hues; yellow-orange pixels fall below 0, and red or magenta pixels (glare, labels, fruit) can exceed 1. These are rare: the 95th percentile never rose above 0.75 in any frame, and the 5th never fell below 0.23.
• NGRDI hit both −1 and +1 at the pixel level, and in every frame the 5th percentile was below zero. At least one plant pixel in twenty was redder than it was green: stems, petioles, edge pixels and shadowed soil caught in the plant mask.
• NDPI was undefined for 0.15% of plant pixels, typically very dark pixels where the denominator approaches zero. NGRDI was undefined for about 11 pixels per frame.
• TGI ran from −15.8 to 43.9 at the pixel level, with a long upper tail. Frame means sat on average 0.53 above frame medians, a sign of right skew driven by a minority of extreme pixels.

Report the median, not the mean. For DGCI, NGRDI and NDPI, frame mean and median differed on average by less than 0.01, but for TGI the mean was 17% higher than the median on average. Medians and percentiles are robust to the glare, stems and edge pixels that every real canopy image contains.

Which greenness index should you use for tomatoes?

For a tomato fertiliser comparison from RGB images, use two indices from different families: DGCI as a stable anchor that is unaffected by canopy size, and TGI (or NGRDI) as the sensitive index that is more likely to show a difference. Report canopy coverage alongside them. Choose the pair before you collect data and keep it for the whole trial.

DGCI's stability cuts both ways. A small fertiliser effect that changes leaf colour only slightly may move DGCI by less than its frame-to-frame noise. That is why a sensitive index belongs next to it, rather than instead of it.

How to compare fertiliser treatments with greenness indices

This protocol turns the lessons of the dataset into a trial design. It works for tomatoes in bags, pots or beds and for most other row crops.

1. Label plants before you film. Tag each plant or bag with its treatment code. Film each treatment block as a separate video, or say the treatment code out loud at the start of each block so the frames can be assigned later.
2. Plan replication at plant level. Each plant is a replicate; each frame is not. Randomise treatments along the row where you can, so a light or slope gradient does not line up with one treatment.
3. Fix the capture. Same height above the canopy, phone pointing straight down, same time of day. Overcast light or full shade is best. Keep your feet and shadow out of the frame: in our video both appear at the edge of several frames, and a shadow that reaches the plants changes their colour.
4. Include a colour reference. Photograph a Greenness Calibrating Plate in the same light at the start of each block, or keep it in frame, so sessions and dates can be compared.
5. Segment plant from background. Compute indices on plant pixels only, and record the plant pixel count and canopy coverage for every frame.
6. Summarise per plant. For each plant, take the median of each index across its frames, plus the 10th and 90th percentiles. Use frames where the whole plant is in view.
7. Compare treatments on plant summaries. Use canopy coverage as a covariate, and report the result both with and without it.
8. Report two indices. One from each family, such as DGCI and TGI, chosen before the trial starts. A treatment difference that both show is a robust result.
9. Validate on a subset. Take SPAD readings or leaf nitrogen analysis on a few plants per treatment to confirm the index ranking matches the physiology.

Seven mistakes that spoil a greenness comparison

1. Treating video frames as replicates. With frame-to-frame correlations of 0.86–0.93, 139 frames behave like a handful of independent observations. Statistics on frames overstate confidence.
2. Filming treatments in a fixed order. If treatment A is always filmed first and treatment C last, any change in light during the walk becomes a "treatment effect".
3. Ignoring canopy size. Bigger plants scored greener on three of the four indices simply by filling more of the frame.
4. Reporting the mean. TGI's mean was 17% above its median, pulled up by a few extreme pixels.
5. Picking the index after seeing the data. With four indices disagreeing, one of them will usually support the result you hoped for. Decide in advance.
6. Comparing absolute values across sessions. A DGCI of 0.53 today and 0.50 next week may reflect cloud cover, not plants. Use a calibration plate and compare within sessions.
7. Reading greenness as nitrogen. Greenness tracks chlorophyll, which also responds to magnesium, iron, water stress, disease and leaf age. Confirm with SPAD or tissue analysis before changing a fertiliser programme.

What this dataset can and cannot tell you

It can tell you:

• how four common RGB greenness indices behave on the same tomato canopy imaged from above;
• how strongly each index responds to canopy coverage rather than colour;
• how far the indices agree, and which ones form a family;
• how much variation sits within a frame compared with along the row;
• why medians, percentiles and plant-level summaries are needed.

It cannot tell you:

• which fertiliser produced the greenest plants. Frames are not labelled by treatment;
• whether the along-row decline in TGI reflects fertiliser, light or plant size. It is consistent with all three;
• how index values relate to leaf nitrogen or chlorophyll content in this crop. That needs SPAD or laboratory validation;
• how the indices behave in other varieties, growth stages, greenhouses or lighting conditions.

Measuring greenness in Petiole Pro

Petiole Pro computes greenness indices from photographs taken with a phone or tablet. Pair it with a Greenness Calibrating Plate and readings taken in different light, on different days or with different devices can be compared on a common reference. The same image also gives leaf area, so canopy size and colour come from one measurement.

To go deeper into greenness measurement:

• Greenness measurement with digital indexes: why greenness is a plant health signal and how DGCI, TGI and AGRI differ.
• Measuring greenness with DGCI: the calibration plate workflow, step by step.
• Relative greenness assessment with RGB indices: control ratios, z-scores and other ways to make a comparison you can defend.
• All greenness and plant nitrogen guides.

Frequently asked questions

References

1. Hunt E. R. Jr., Doraiswamy P. C., McMurtrey J. E., Daughtry C. S. T., Perry E. M., Akhmedov B. (2013). A visible band index for remote sensing leaf chlorophyll content at the canopy scale. International Journal of Applied Earth Observation and Geoinformation 21, 103–112. doi:10.1016/j.jag.2012.07.020
2. Karcher D. E., Richardson M. D. (2003). Quantifying turfgrass color using digital image analysis. Crop Science 43(3), 943–951. doi:10.2135/cropsci2003.9430
3. Tucker C. J. (1979). Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of Environment 8(2), 127–150. doi:10.1016/0034-4257(79)90013-0