AI Assessment of Necrotic Leaf Area Damage: From Detecting Disease to Measuring It

What necrotic leaf area damage is — and why measure it

Necrosis is dead plant tissue. On a leaf it shows up as brown, tan, grey, or dry patches where cells have collapsed — often the end stage of disease, sunburn, nutrient deficiency, or environmental stress. Necrotic leaf area damage is simply the proportion of a leaf given over to that dead tissue. It is one of the most common yet most laborious things a plant scientist, agronomist, or breeder is asked to quantify.

Why bother turning it into a number? Because a subjective glance cannot compare thousands of leaves, cannot be reproduced by a second person, and cannot show a trend over time. Quantifying necrosis converts a visual impression into data you can benchmark across treatments, track across a season, and feed into a decision.

Before you measure necrosis, you have to decide what counts as necrosis. Not every yellow, pale, or brownish area is dead tissue yet — some lesions are only discoloured. Defining necrotic tissue is the part where human knowledge is still essential.

Detecting disease vs. measuring damage

Most plant-disease AI answers a yes/no question: is the disease present? That is useful for scouting and early warning. But a grower deciding whether to spray, a breeder ranking cultivars for resistance, and a researcher proving a treatment works all need something more precise — the severity, not just the diagnosis.

This is the shift from detection to quantification: from "this leaf has rust" to "this leaf is 48.7% necrotic, and last week it was 32%." The direction and speed of that number is what tells you whether disease pressure is building and whether your intervention is holding.

How AI quantifies necrosis from a photo

The workflow behind the number is, in simplified terms, three steps: separate the leaf from its background, classify each pixel of the leaf as healthy or damaged, then add up the areas. The first step is segmentation — the same operation used in ordinary leaf-area measurement, where the algorithm finds the border between leaf and everything else. Draw a second border inside the leaf and the algorithm sees two objects: a healthy region and a damaged one.

The second step is where colour indices come in. Healthy green tissue reflects light differently from yellow, brown, dry, or dead tissue, so the algorithm can sort pixels by their colour signal. Research on coffee leaf rust (Hemileia vastatrix) has tested a whole family of RGB indices for exactly this task.

RGB indices do not "see" the biology. They classify colour patterns. This is powerful — but it is not magic, which is exactly why a fixed definition of necrosis and a human check still matter.

RGB colour indices: NGRDI, AGRI and custom

There are dozens of vegetation indices, and any of them can be applied to the same image. In practice only a few are useful for a given task — the rest may return an almost-black or almost-white map that separates nothing. For necrotic-area work on coffee, two earned their place:

• NGRDI (Normalized Green–Red Difference Index): a colour/vegetation index that reacts to the loss of green relative to red. Positive values indicate chlorophyll is present; negative values indicate chlorophyll loss — the signature of dying tissue.
• AGRI-style index: a second colour/vegetation index, sensitive to slightly different colour behaviour, used alongside NGRDI as a cross-check.
• A custom index: indices are a creative tool — anyone can define one to sharpen the segmentation of a particular kind of lesion. We used a customised index at the final stage to clean up the damaged-area boundary.

Rendered as a heatmap, the index makes the damage legible even to a non-expert. On these coffee leaves the healthy blade glows red-orange while necrotic and rust-affected zones fall away into blue and violet. Note the display range printed on each map (for example disp −0.300 .. 0.500): the palette is stretched to the values in that specific image, which is why two leaves can look different unless the ranges are standardised.

Different indices may give different necrosis estimates — and that is not a bug. In our report AGRI estimated a larger damaged area than NGRDI, because each index is sensitive to slightly different colour behaviour. A combined result reduces dependence on any single index.

From damage map to healthy-vs-necrotic percentage

Once pixels are classified, the arithmetic is simple: the necrotic area divided by the total leaf area gives the damage percentage, and its complement is the healthy fraction. Colour the two classes back onto the leaf and you get an interpretable result rather than a raw CSV of ten-figure numbers.

That single percentage is exactly what the researcher who first sent us this leaf asked for. It is also the kind of number that, in the future, an AI agent could act on automatically — reading "healthy 85.5%" and starting a decision cycle. For now, humans do that reasoning; the visual assessment is the foundational data the agents of tomorrow will be built on.

Measuring necrosis on your phone with Petiole Pro

You do not need a lab to get a first number. The Petiole Pro mobile app already measures leaf area on-device, and the same tool can separate a damaged region from a healthy one. The manual technique is deliberately low-tech: photograph the leaf next to a calibration plate, then draw a line around the necrotic part.

Analyse the leaf and the app returns each region separately, in real units. Tap the healthy region and you get its area and perimeter; tap the necrotic region and you get its own. Subtract or divide, and you have your percentage backed by absolute cm².

The same method works for any kind of damage, not only disease. Sun and heat injury — thermal burn — can be quantified in exactly the same way: measure the whole leaf, then the burned region, and report the percentage.

Eight rules for trustworthy necrosis measurements

Getting a number is easy; getting a number you can defend takes discipline. These are the practical rules distilled from working through real leaves.

1. Percentages are fine for production. At production level, assessing damage in percent is enough. Two indices can be sufficient to understand the dynamics, combined into a single healthy/damaged indicator for decision-making.
2. One number is never enough — measure on a timeframe. Recalculate every 5–7 days, depending on crop, disease pressure, and conditions. Two or more measurements in a row are what let you compare and see the direction.
3. For science, use both relative and absolute measurements. Relative tells you the percentage (e.g. 32% damaged); absolute tells you the real area in cm². The two together are a different level of data quality.
4. For absolute numbers, calibration is mandatory. A calibration plate — or any calibration object in ImageJ, MATLAB, or another tool — is what converts pixels to cm². Without it you can only claim percentages.
5. Define necrosis before you measure it. Some lesions are discoloured but not yet dead. Decide what you call necrotic tissue first, so a yellow or pale area is not silently counted as necrosis.
6. Remember what RGB indices actually do. They classify colour patterns, not biology. Healthy green behaves differently from yellow, brown, or dry tissue — powerful, but not a diagnosis.
7. Expect indices to disagree — and use it. A stricter index and a more conservative one give different estimates; combining them reduces dependence on any single method.
8. The real deliverable is a repeatable workflow. Same crop, same lighting logic, same calibration approach, same timeframe, and the same definition of necrosis. Only then do the numbers become genuinely comparable in the natural environment.

Watch the full walkthrough

The workflow above — segmentation, indices, the combined percentage, and the on-phone demo — was walked through live. The recording covers the coffee-leaf-rust examples and the reasoning behind each choice.