Leaf area measurement
Potato leaf area measurement is the quantification of the photosynthetic surface of a potato (Solanum tuberosum) plant from an image, in real units, without cutting the leaf off the plant. On potato it is harder than on most crops for one anatomical reason: the potato leaf is imparipinnate compound — a single leaf is a rachis carrying a large terminal leaflet, three to four pairs of primary lateral leaflets, and a scatter of smaller interstitial leaflets between them. What you call “one leaf” decides your number before the camera is even raised.
In Petiole R&D Lab experiment 006, run on 20 August 2026, 53 potato compound leaves were photographed by a Petiole app user and processed automatically on the Petiole Pro Web platform. The whole compound leaf was treated as the measurement unit. The result: a mean total leaf area of 43.84 cm², a median of 38.07 cm², and a range running from 13.63 cm² to 89.18 cm² — a 6.5-fold spread inside a single crop, at a single sampling. Mean detection confidence across all 53 leaves was 0.982.
This long-read explains what that variation is made of, why it is a biological result rather than a measurement failure, how the compound-leaf structure changes every step of the protocol, how many leaves you need before a potato treatment mean means anything, and how to run the whole thing on a phone in a wet, muddy, wind-moving canopy. If you want the general method first, start with our guide to photographing leaves for accurate leaf area measurement in the field and come back here for the potato specifics.
- Declare your unit before you shoot. Terminal leaflet, whole compound leaf, or whole plant are three different measurements. Mixing them inside one dataset is the single most common way potato leaf-area work goes wrong.
- Potato leaf size varies enormously within one plant. Experiment 006 recorded 13.63 cm² to 89.18 cm² across 53 compound leaves — mostly node position and leaf age, not measurement error.
- Use the median, report the spread. Mean 43.84 cm² against median 38.07 cm² is a right-skewed distribution: a minority of large, fully expanded leaves pulls the average above the typical leaf.
- A coefficient of variation near 43% sets your sample size. Detecting a 10% treatment difference at 80% power needs roughly 300 leaves per group; a 20% difference needs roughly 75. Three leaves per plot is not a measurement, it is a mood.
- Automated segmentation was reliable here. Mean detection confidence 0.982, median 0.985, minimum 0.967 — every one of the 53 leaves cleared 0.96, which is what a clean black-background capture on a compound leaf looks like.
- Non-destructive is the whole point. Potato canopy duration drives tuber bulking as much as canopy size, and duration can only be measured by returning to the same plants.
Why potato is not like other crops
Most leaf-area workflows were designed around a simple leaf: one blade, one petiole, one outline. Cotton, basil, soybean, mangrove species — you flatten the blade, you photograph it, the software traces a single closed contour and returns an area. Our independent validation of Petiole Pro against the LI-COR 3100 on cotton is exactly that kind of study, and it reached correlations up to r = 0.988 because the object being measured was unambiguous.
Potato is not that object. A mature potato leaf is a branched structure. Along its rachis sit paired primary leaflets that decrease in size towards the base, a single dominant terminal leaflet at the tip, and — depending on cultivar — secondary and tertiary interstitial leaflets tucked between the primaries. Leaflet number, leaflet shape and the density of interstitials are cultivar characteristics; they are part of how potato varieties are described botanically in the first place.
That structure creates four measurement problems at once, and every one of them is specific to compound-leaved crops.
| Problem | What goes wrong | What fixes it |
|---|---|---|
| Ambiguous unit | “Leaf area = 21 cm²” could be one leaflet or one whole leaf. The two differ by a factor of five or more. | Write the unit into the method before sampling and never change it. |
| Non-coplanar leaflets | Leaflets sit at different angles to the rachis, so a photograph captures a projection, not the true surface — each tilted leaflet is silently undersized. | Flatten the whole leaf against the plate before shooting; accept that projected area is the definition. |
| Self-occlusion | Overlapping leaflets hide tissue from the camera. Shared pixels are counted once, so the total is too low. | Separate leaflets by hand until none touch, then shoot. |
| Rachis and petiolules | Green stem tissue between leaflets is photosynthetic but is not lamina. Including it inflates area; excluding it is harder than it sounds. | Pick one convention, state it, and keep every leaf in the study on the same side of the line. |
None of these are software problems. They are definition problems, and software cannot fix a definition you never made. The good news is that once the definition is fixed, computer vision handles a potato leaf about as well as it handles a simple one — as the confidence numbers below show.
The unit problem: leaflet, leaf, or plant?
There are three defensible units for potato leaf area, each answering a different question. What is not defensible is drifting between them within a dataset, or failing to say which one you used.
- The individual leaflet. Highest throughput per photograph, cleanest segmentation, and the easiest thing to standardise — the terminal leaflet of the youngest fully expanded leaf is a well-defined, repeatable object. It is the right unit for leaf-level physiology, greenness sampling and nutrient work. It is the wrong unit for anything scaling to canopy, because leaflet count varies between cultivars and along the stem.
- The whole compound leaf. The unit used in experiment 006. It captures the entire leaf as the plant builds it, including the small interstitial leaflets that a leaflet-only protocol throws away, and it is the unit that scales to the plant by simple multiplication with leaf count. It costs more handling time per sample and demands that leaflets be separated before the shot.
- The whole plant or canopy. The unit agronomy actually cares about, because tuber bulking is driven by intercepted light. It is reachable either destructively — strip every leaf and measure them all — or by combining a measured mean leaf area with a counted leaf number, which is the non-destructive route this article is built around.
Which should you choose? If the question ends in “per plant”, “per m²” or “LAI”, use the whole compound leaf. If the question ends in “per unit leaf area” — nitrogen concentration, greenness index, specific leaf weight, disease severity — use the leaflet, because the leaflet is where the physiology is uniform. Never average leaflet-based and leaf-based numbers into one column.
Experiment 006: 53 leaves, one afternoon
The dataset behind this article was collected by a Petiole mobile app user in the field and processed automatically on the Petiole Pro Web platform. Fifty-three potato compound leaves were photographed, segmented and measured under a single parameter set, with no per-image adjustment and no manual retouching of outlines.
Sorting the grid by area makes the structure of the sample visible before any statistics are computed. The largest leaves are broad, dark and symmetrical, with wide terminal leaflets and full complements of laterals. Towards the bottom the leaves are smaller, paler, more chlorotic at the margins, and several carry visible necrotic patches and feeding damage. That gradient is the point: a potato canopy at any given moment contains leaves at every stage of that sequence simultaneously.
| Statistic | Value | What it tells you |
|---|---|---|
| Leaves measured (n) | 53 | A real sample, not a demonstration — but see the sample-size section below. |
| Mean total leaf area | 43.84 cm² | The number most people quote. It is pulled upward by the large leaves. |
| Median | 38.07 cm² | The typical leaf. 13% below the mean, which is the skew made numeric. |
| Minimum | 13.63 cm² | Young, senescing or damaged leaves at the bottom of the canopy. |
| 25th percentile | 31.16 cm² | A quarter of all leaves were smaller than this. |
| 75th percentile | 57.08 cm² | A quarter were larger. The interquartile range is 25.92 cm². |
| Maximum | 89.18 cm² | 6.5× the smallest leaf in the same sample. |
| Standard deviation | 19.04 cm² | A coefficient of variation of about 43% — high, and the driver of everything downstream. |
Reading the distribution, not the average
A mean of 43.84 cm² is a fact about this sample. On its own it is close to useless, because it hides the thing that actually determines what you can conclude: how widely potato leaf size varies within a single canopy at a single moment.
Three features of this distribution matter for anyone designing a potato experiment.
It is right-skewed. The mean sits 15% above the median. In a canopy, leaf area accumulates towards the large end: fully expanded mid-stem leaves are several times the size of the young leaves at the apex and the senescing ones at the base. Any statistic that assumes symmetry — a mean with a standard deviation, a t-test on raw areas from a small sample — is working against the shape of the data. Report the median with the interquartile range alongside the mean, and consider a log transform before parametric testing.
The spread is enormous. A standard deviation of 19.04 cm² against a mean of 43.84 cm² is a coefficient of variation of about 43%. For comparison, that is far larger than the measurement error of any reasonable leaf-area method. The variation you are seeing is the plant, not the instrument — which is good news for the instrument and bad news for anyone hoping to run a treatment comparison on five leaves.
Perimeter behaves differently from area. This is a compound-leaf effect worth understanding. On a simple leaf, perimeter and area move together in a fairly tight relationship. On a potato leaf, perimeter counts the outline of every leaflet separately, so a leaf with many small interstitials has a much longer perimeter than a leaf of identical area with fewer, larger leaflets. That makes the area-to-perimeter ratio a cheap, useful descriptor of leaf dissection — and it makes perimeter a poor proxy for size on this crop.
On a compound leaf, area tells you how much the plant built. Perimeter tells you how it chose to divide it up.
What a detection confidence of 0.982 means
Every automated measurement in experiment 006 carried a detection confidence score — the model's own assessment of how cleanly it separated leaf tissue from everything else. Across the 53 leaves the mean was 0.98226, the median 0.984515, the 25th percentile 0.98007 and the minimum 0.967025.
What it does mean: the segmentation was consistent. There was no subset of images where the model struggled — no partial leaves, no background bleeding into the contour, no leaflets dropped. The minimum being 0.967 rather than, say, 0.61 is the important part: in a batch, the worst confidence tells you far more than the average, because a single badly segmented leaf can move a small treatment mean noticeably.
What it does not mean: confidence is not accuracy against a reference instrument. A model can be entirely confident about a contour it has drawn around a leaf whose leaflets were overlapping, or one photographed at an angle, or one whose calibration reference was partly out of frame. Confidence measures whether the pipeline found a clean object. It cannot measure whether you gave it the right object. That is what the protocol below is for, and it is why validation against a laboratory meter — as in the peer-reviewed cotton study — is a separate exercise from per-image confidence scoring.
Treat detection confidence as a quality gate, not a result. Set a floor before you analyse — 0.95 is a reasonable starting point for clean black-background captures — apply it identically to every treatment group, and report how many leaves each group lost. A group that loses far more images than the others has a photography problem, not a biological one.
How many potato leaves do you actually need?
This is where the 43% coefficient of variation stops being a statistic and starts being a budget. Sample size on potato leaf area is not a matter of taste; it falls straight out of the spread measured above.
Start with the simpler question: how precisely does a given sample pin down the mean? With a standard deviation of 19.04 cm², the half-width of a 95% confidence interval around the mean is 1.96 × 19.04 / √n.
| Leaves measured | 95% CI half-width | As a share of the 43.84 cm² mean |
|---|---|---|
| 5 | ±16.7 cm² | ±38% |
| 10 | ±11.8 cm² | ±27% |
| 20 | ±8.3 cm² | ±19% |
| 53 (experiment 006) | ±5.1 cm² | ±12% |
| 100 | ±3.7 cm² | ±8.5% |
| 200 | ±2.6 cm² | ±6.0% |
Read the top row carefully. A five-leaf sample from this population produces a mean whose 95% interval spans nearly 80% of its own value. It is not a weak measurement — it is not a measurement. The 53 leaves of experiment 006 place the mean somewhere between roughly 38.7 and 49.0 cm², which is useful for describing a canopy and still too coarse for detecting a modest treatment effect.
For comparing two groups, the arithmetic is harsher. At 80% power and a 5% significance level, the leaves needed per group to detect a difference of a given size, at a coefficient of variation of 43%, are approximately:
| Difference you want to detect | In cm² (from a 43.84 cm² baseline) | Leaves per group |
|---|---|---|
| 5% | 2.2 cm² | ~1,180 |
| 10% | 4.4 cm² | ~300 |
| 15% | 6.6 cm² | ~130 |
| 20% | 8.8 cm² | ~75 |
| 30% | 13.2 cm² | ~35 |
| 50% | 21.9 cm² | ~12 |
The practical reading: subtle potato leaf-area effects need hundreds of leaves per treatment, which is precisely why automated, batch-processed image measurement changes what is feasible. Photographing 300 leaves is an afternoon's work. Running 300 leaves through a laboratory meter — after detaching all 300 from the plants — is a different kind of project entirely, and it ends the time series.
Two ways to shrink the requirement honestly. Standardise leaf position: sampling only the youngest fully expanded leaf on every plant removes most of the node-driven variance, cutting the effective CV substantially. Use a paired design: measuring the same tagged leaves before and after a treatment, or the same plants across dates, removes between-leaf variation from the comparison entirely — and only a non-destructive method lets you do it.
From leaf area to leaf area index
Leaf area per leaf is a phenotyping number. What agronomy wants is leaf area index (LAI) — square metres of leaf per square metre of ground — because that is what determines how much of the incoming radiation the crop intercepts, and interception drives tuber bulking.
The bridge is arithmetic:
with leaf area in cm² and the 10,000 converting cm² to m².
Worked with experiment 006's mean: a plant carrying 40 compound leaves at 43.84 cm² each has about 1,754 cm² — roughly 0.18 m² — of leaf. At a typical field density of about four plants per m², that gives an LAI of approximately 0.70. Later in the season, with 60 leaves per plant averaging nearer the 89 cm² top of this sample's range, the same arithmetic gives an LAI above 2.
That trajectory is the measurement. A single LAI value is a snapshot; the curve of LAI over the season — how fast the canopy closes, how high it peaks, and crucially how long it stays there before senescence — is what separates a high-yielding potato crop from an average one. Canopy duration is only measurable by returning to the same plants, which is the argument for non-destructive measurement in one sentence.
Two cautions on this arithmetic. Leaves per plant has to be counted, and counting it accurately in a closed potato canopy is genuinely hard — expect it to be the larger source of error, not the leaf area. And the mean leaf area used must come from a properly stratified sample across the stem, not from whichever leaves were easiest to reach, or the multiplication propagates a sampling bias straight into the LAI.
What potato leaf area is actually for
Leaf area is rarely the end goal. It is an intermediate variable that several very different questions pass through.
| Question | What leaf area contributes | Unit to measure |
|---|---|---|
| Nitrogen management | Nitrogen supply drives canopy expansion and delays senescence. Leaf area growth rate responds to N status before yield does. | Whole compound leaf, weekly, fixed node |
| Late blight and early blight severity | Blight is scored as loss of functional green canopy. Measuring healthy area and lesion area on the same image turns a visual score into a number. | Leaflet, with damage segmentation |
| Colorado potato beetle defoliation | Action thresholds are expressed as percentage defoliation, which is a leaf-area ratio measured against an undamaged baseline. | Whole compound leaf, paired over time |
| Drought and heat stress | Leaf expansion is one of the first processes to shut down under water deficit, well before visible wilting or measurable yield loss. | Whole compound leaf, same tagged leaves |
| Cultivar and clone screening | Maturity class is largely a statement about canopy duration. Leaf-area curves separate early from late material long before harvest. | Whole compound leaf, many plants per entry |
| Growth analysis and modelling | Relative growth rate, net assimilation rate and specific leaf area all require leaf area as a denominator. | Whichever unit the model was parameterised on |
The blight and beetle rows deserve a note, because they change what you are measuring. A leaf with feeding holes and necrotic lesions returns a lower area than the same leaf undamaged — the missing tissue is genuinely missing from the projection. That is correct behaviour if your question is “how much functional canopy is left”, and it is a confound if your question is “how large did this leaf grow”. Several leaves in the lower half of the experiment 006 grid carry exactly this damage. If you need to separate the two questions, pair area measurement with explicit damage assessment — our guide to AI-based necrotic leaf area assessment covers that workflow.
A field protocol for potato
Everything above collapses into a repeatable sequence. This is the protocol experiment 006 followed, with the potato-specific decisions made explicit.
- Declare the unit. Whole compound leaf or individual leaflet. Write it in the method file before you go to the field. Do not change it mid-season, mid-trial or mid-anything.
- Declare the leaf position. The youngest fully expanded leaf is the standard choice, because it is identifiable on every plant regardless of size or growth stage. Fixing position is the cheapest variance reduction available to you.
- Bring the calibration reference. Potato leaves are photographed at close range in a canopy where working distance varies constantly. The plate is what converts pixels to cm² — why the calibration plate is not optional is worth reading if you are tempted to skip it.
- Detach or support, but decide which. Petiole Pro measures leaves in place on the plate. Whether you detach the leaf or bring the plate to it, do it the same way every time — a supported in-place leaf and a detached flattened leaf lie differently.
- Separate the leaflets. The potato-specific step. Spread the leaf so that no leaflet overlaps another and none folds back on itself. Overlap is the largest avoidable source of underestimation on this crop.
- Flatten. Press cupped or curled leaflets flat against the plate. Potato leaflets cup under water stress and heat, and a cupped leaflet loses projected area that has nothing to do with how large it grew.
- Shoot perpendicular, in diffuse light. Camera parallel to the plate, whole plate in frame, shadow or overcast rather than direct sun. Potato foliage is pubescent and throws specular highlights that eat into the segmented edge.
- Sample for the distribution. Use the sample-size table above. Replicate across plants and plots, not across leaves of one plant — leaves on the same stem are not independent observations.
- Batch process under one parameter set. Every image in the study through the same pipeline, with no per-image tuning. This is what makes the comparison between groups fair.
- Apply the confidence gate. Filter on detection confidence at a threshold declared in advance, and record the exclusions per group.
- Report median, IQR, n and unit. Not just the mean. A potato leaf-area result without the spread and the unit definition cannot be interpreted by anyone else, including you in six months.
Seven mistakes specific to potato
- Calling a leaflet a leaf. The single biggest error. A terminal leaflet might be 15 cm² where the whole leaf is 45 cm². Datasets that mix the two are unrecoverable after the fact, because nothing in the numbers tells you which rows are which.
- Sampling whichever leaves are convenient. The 13.63–89.18 cm² range in experiment 006 is mostly node position and leaf age. Sampling the outer, accessible leaves of a canopy systematically oversamples one part of that range.
- Letting leaflets overlap. Shared pixels are counted once. A leaf photographed with two leaflets touching is measured smaller than it is, and the error is invisible in the output — the segmentation looks perfect.
- Ignoring leaflet cupping. Under heat or water stress, potato leaflets curl upward. If you photograph them curled, you record a stress response as a size difference, and the two are not the same variable.
- Being inconsistent about the rachis. Whether the green rachis and petiolules are inside the contour changes total area by a few percent — small, but the same order as the treatment effects people try to detect. Consistency matters more than which choice you make.
- Averaging damaged and undamaged leaves without saying so. Blight lesions and beetle feeding remove real area. That belongs in a canopy-loss analysis, not silently inside a growth mean.
- Shooting in direct sunlight. Potato foliage is hairy and glossy in strong light. Specular highlights push the segmented edge inward, and the resulting bias is systematic across the whole session — which makes it far more dangerous than random noise.
Running it in Petiole Pro
The workflow in experiment 006 was deliberately ordinary: a user photographed leaves in the field on a phone, and the images were processed as a batch on the Petiole Pro Web platform. No image was hand-corrected. What made the output usable was not exotic tooling but three properties of the pipeline.
- Real units from a reference in frame. The calibration plate visible in the field photograph at the top of this article carries the fiducial markers that fix scale, so a leaf photographed at arm's length and one photographed close in return comparable areas.
- A quality score attached to every measurement. Detection confidence travels with each row, so the quality gate is a filter on the dataset rather than a judgement call on individual images.
- Batch processing with one parameter set. Fifty-three leaves — or five hundred — through the same configuration, exported as labelled, filterable rows with area, perimeter and confidence per leaf.
The same capture also supports colour work: if canopy nitrogen status is part of the question, the greenness indexes computed from the same photographs give a second variable at no extra fieldwork cost. Our field guide to measuring greenness with DGCI covers that side of the workflow.
Honest limits
Four things this method does not do, stated plainly.
It measures projected area, not true surface area. Every image-based method does. For a flattened leaf on a plate the two are close; for a leaf with permanently curled leaflets they are not, and the difference is real rather than a bug. If your question is about true surface area — boundary-layer physics, for instance — projection is the wrong measurement.
It does not count leaves for you. Scaling to LAI needs leaves per plant, and in a closed potato canopy that count is the harder measurement of the two. Budget error there, not in the leaf area.
Confidence is not accuracy. Experiment 006 demonstrates consistent segmentation across 53 leaves. It is not a validation against a laboratory leaf-area meter on potato; that is a separate study with a separate design, of the kind published on cotton.
One session is one session. These 53 leaves came from one collection on one date. They describe what potato leaf-size variation looks like in that canopy — enough to size a sample and design a protocol, not enough to serve as a reference value for the crop. Your own baseline is the one worth measuring.
Petiole Pro for potato phenotyping
Measure the canopy you actually have
Petiole Pro measures leaf area, perimeter and greenness from an ordinary smartphone photograph, attaches a detection confidence score to every reading, and batch-processes whole folders into labelled, filterable, export-ready rows. The Petiole Pro calibration plates are the reference object that turns pixels into cm² in a field where working distance never stays still.
Whether you are timing a nitrogen application, scoring blight progression, screening clones for canopy duration or tracking how a drought treatment slowed leaf expansion, the shape of the answer is the same: this group against that group, same unit, same protocol, same pipeline, with the exclusions declared and the spread reported.
Frequently asked questions
What is the average leaf area of a potato leaf?
In Petiole R&D Lab experiment 006, across 53 potato compound leaves measured on 20 August 2026, the mean total leaf area was 43.84 cm² and the median 38.07 cm², with a range from 13.63 cm² to 89.18 cm² and a standard deviation of 19.04 cm². Those figures describe whole compound leaves — rachis, terminal leaflet, primary laterals and interstitials together — from one canopy on one date. They are useful for sizing a sample and setting expectations, not as a reference value for the crop: potato leaf area depends heavily on cultivar, node position, leaf age, growth stage, nutrition and stress history.
Should I measure a potato leaflet or the whole compound leaf?
It depends on the question. Use the whole compound leaf when the result scales to the plant or canopy — leaf area per plant, leaf area index, growth analysis — because leaflet counts vary between cultivars and along the stem. Use the individual leaflet, typically the terminal leaflet of the youngest fully expanded leaf, when the result is expressed per unit leaf area: nitrogen concentration, greenness index, specific leaf weight or disease severity. What matters most is declaring the unit before sampling and never mixing the two inside one dataset.
Why does potato leaf area vary so much within one plant?
Because a potato stem carries leaves at every stage of development at once. Leaves at the apex are still expanding, mid-stem leaves are fully expanded and at maximum area, and basal leaves are senescing or have been damaged by disease and feeding. Experiment 006 recorded a 6.5-fold range from 13.63 cm² to 89.18 cm² in a single sample, giving a coefficient of variation near 43%. Sampling a fixed leaf position, usually the youngest fully expanded leaf, removes most of this variance and is the cheapest improvement available to any potato leaf-area protocol.
How many potato leaves should I measure per treatment?
With the roughly 43% coefficient of variation observed in experiment 006, detecting a 10% difference between two groups at 80% power and a 5% significance level needs on the order of 300 leaves per group; a 20% difference needs about 75, and a 30% difference about 35. A five-leaf sample gives a 95% confidence interval on the mean of roughly plus or minus 38%, which is too wide to support any comparison. Standardising leaf position or using a paired design on tagged leaves reduces these requirements substantially.
How do you measure potato leaf area without destroying the leaf?
Photograph the leaf against a calibration plate that carries a known scale reference, and let computer vision segment the leaf from the background and convert pixels to square centimetres. Nothing is detached and nothing is damaged, so the same tagged leaf can be measured again a week later. This is what makes canopy duration measurable, and canopy duration matters as much as canopy size for potato tuber bulking, because both feed into the total radiation the crop intercepts across the season.
What does the detection confidence score tell me?
It reports how cleanly the model separated leaf tissue from the background in that image, not how accurate the measurement is against a laboratory reference. In experiment 006 the mean was 0.982, the median 0.985 and the minimum 0.967, so every leaf in the batch segmented cleanly. Use it as a quality gate: set a floor before analysis, for example 0.95 for black-background captures, apply it identically to every treatment group, and report how many leaves each group lost. In a batch the minimum confidence is more informative than the mean.
How do I convert potato leaf area to leaf area index?
Multiply the mean leaf area per leaf in cm² by the number of leaves per plant and by the number of plants per square metre, then divide by 10,000 to convert cm² to m². Using experiment 006's mean, a plant with 40 compound leaves of 43.84 cm² carries about 0.18 m² of leaf, which at four plants per m² is an LAI of roughly 0.70. The larger source of error in this calculation is usually the leaf count in a closed canopy, not the leaf area, and the mean leaf area must come from a sample stratified across the stem rather than from the most accessible leaves.
Do blight lesions and beetle feeding affect the measurement?
Yes, and correctly so. Missing tissue is genuinely missing from the projection, so a damaged leaf returns a smaller area than the same leaf undamaged. That is the right answer if you are quantifying remaining functional canopy, and a confound if you are quantifying how large the leaf grew. Decide which question you are asking, declare it, and if you need both, pair area measurement with explicit damage segmentation rather than letting damaged leaves sit unmarked inside a growth mean.
Why is perimeter a poor proxy for potato leaf size?
Because on a compound leaf the perimeter traces the outline of every leaflet separately. Two leaves of identical area can have very different perimeters depending on how many leaflets they carry and how dissected those leaflets are, which is exactly why the perimeter histogram from experiment 006 is broader and lumpier than the area histogram. Perimeter is still useful — the area-to-perimeter ratio is a cheap descriptor of leaf dissection — but it should not be used as a size measure on this crop.
Do I really need a calibration plate for potato?
Yes, and arguably more than for most crops. Potato leaves are photographed inside a canopy where working distance changes with every plant, and without a known reference in frame there is no way to convert pixels into square centimetres consistently between shots. The plate also provides the colour reference that keeps greenness values comparable if you compute them from the same photograph, and gives the segmentation a high-contrast background to work against.
The useful question is never “how big is a potato leaf”. It is “which unit did you measure, at which node, on how many plants, and what did the spread look like”.