A Phone App That Reads Strawberry Leaves Matched a Standard Harvest, but It Used More Than Three Times the Fertilizer

On terraced hillsides in Meghalaya, in northeast India, a team of agricultural researchers spent a strawberry season testing a simple idea. A grower takes photos of leaves with an ordinary Android phone. An app estimates how much nitrogen the plants are carrying, works out how much they are missing, and tells the grower how many minutes to run fertilizer through the drip line.

The app, called NitroMate, was built at the ICAR Research Complex for North Eastern Hill Region in Umiam, part of the Indian Council of Agricultural Research. The results, published in the journal Results in Engineering, are a useful lesson for anyone who has ever trusted a gadget or an app to tell them how much to feed a plant. The phone did a reasonable job of spotting hungry plants. The harvest came out close to what a standard fertilizer schedule produced. But the app recommended far more nitrogen than the plants needed, more than three and a half times as much over the season.

The paper is titled “Smart fertigation: A deep learning-based hand-held decision support system for plant-responsive nitrogen management in strawberry production.” Its authors include Naseeb Singh, Kethavath Ajaykumar, Anuradha Sahu, R. Thangavel, V. K. Verma, B. U. Choudhury and G. Kadirvel. The details below come from a summary of the study published Oct. 26 by the science news site Bioengineer.org, along with the journal’s publication record.

Teaching a phone to read a leaf

Nitrogen is the nutrient strawberry growers worry about most. It is part of chlorophyll and the proteins that drive growth. Too little stunts the plants and cuts the harvest. Too much pushes them into leafy growth at the expense of fruit, and the excess can wash into groundwater.

The standard way to measure nitrogen in leaves is a lab test that is accurate but slow. Handheld chlorophyll meters are faster, but their readings depend on which leaf you clip and need to be calibrated for each crop. Satellite and spectral sensors cover more ground, but small farms on steep terraces rarely have access to them.

To build their training data, the researchers first calibrated a handheld chlorophyll meter against lab nitrogen tests on 30 plant samples. That let them attach a nitrogen estimate to each leaf photo. They took about 3,400 leaf images of the cultivar Sweet Sensation between 45 and 100 days after transplanting, kept 3,000 that passed quality checks, and sorted them into eight nitrogen levels, from below 2.25 percent to above 3.75 percent.

Instead of judging one leaf, their model looks at five leaf photos from the same plant together and makes a single call for the whole plant. That design reflects something gardeners already know: one odd-looking leaf does not tell you much.

How well it spotted hungry plants

Because the real question in the field is yes or no, whether a plant needs more nitrogen, the team tested the model on 75 plants it had never seen. It was right about 79 percent of the time. Of the 43 plants that truly needed nitrogen, it flagged 40 and missed three. It also raised 13 false alarms on plants that did not need any more.

In other words, the model leaned toward saying “feed it.” For a farmer worried about losing yield, catching nearly every hungry plant is useful. The cost of that caution shows up later in the story.

The researchers then shrank the model to run on a phone. On the device it reached about 77 percent accuracy, only slightly below the full version, and it processed a set of five leaf photos in about two seconds.

From a photo to minutes on the drip line

NitroMate’s workflow is meant to be simple. A grower photographs leaves from 10 to 12 plants on a terrace, and the app averages their nitrogen estimates. The grower then enters three numbers: how fast the drip emitters run, the crop’s growth stage, and how much urea fertilizer is mixed into the irrigation water.

The app compares each plant’s estimated nitrogen with a fixed target of 4.0 percent, the top of the published sufficiency range for strawberry leaves, and scales the gap by the size of the plant at that growth stage. It converts that shortfall into grams of urea, then into a volume of water, and finally into a run time in minutes.

In the field trial, nine terraces were split into three groups of three. One group got NitroMate’s recommendations. One got a fixed seasonal schedule of 80 kilograms of nitrogen per hectare. One got no nitrogen at all. Fertilizer went on once a week for 10 weeks, and the AI group’s leaves were photographed right before each application. The app’s recommended run times ranged from about 11 minutes to 50 minutes and changed from terrace to terrace and week to week, so it was responding to the plants rather than repeating one dose.

The harvest, and the catch

The yields were close. The AI-guided terraces averaged about 20,745 kilograms of fruit per hectare, compared with about 20,909 for the fixed schedule, a difference of less than 1 percent. The no-nitrogen terraces produced about 17,477. Average berry weight was 24.3 grams with the app, 24.6 grams with the fixed schedule and 21.9 grams with no nitrogen.

The researchers are careful about what that means. The statistical tests found no significant difference between treatments. The trial was not designed to prove the two methods were equal, the treatments were assigned to neighboring groups of terraces on a slope rather than at random, and the study could only have detected very large differences. The similar yields are an observation, not proof.

The real problem was fertilizer. Over the season, the AI-guided terraces received about 282 kilograms of nitrogen per hectare, more than three and a half times the 80 kilograms in the fixed schedule. Measured another way, the fixed schedule produced about 262 kilograms of fruit for every kilogram of nitrogen applied. The app’s approach produced about 73.

The researchers traced the overdose to the dosing math, not the photo reading. Each week, the app calculated a fresh shortfall from the plants’ current state without subtracting the fertilizer it had already recommended in earlier weeks. And because every calculation aimed at the fixed 4.0 percent target at the top of the healthy range, it kept recommending more. The team named that recommendation step as the main thing to fix next. They also noted that they did not measure long-term soil effects, nitrogen losses, costs or environmental effects, and that the system needs testing across more seasons, places and varieties before wider use.

What it means for the rest of us

This is a small study on one cultivar in one season, and NitroMate is a research tool, not something you can download for a backyard strawberry bed. Still, the result is useful because it separates two jobs that plant apps often blur together.

The first job is diagnosis: is this plant short on something? Here, a phone camera and a well-trained model did reasonably well, and the study’s summary notes that this kind of setup could matter on small terraced farms where expensive sensors and machinery are not practical. The second job is the prescription: how much should I add? That is where the app went wrong, and it went wrong in a way any gardener can recognize. It aimed for the top of the healthy range and did not account for what had already gone into the soil.

The same caution applies at home. A fertilizer recommendation is only as good as its assumptions about what the plant already has, and more nitrogen did not mean more fruit here. The fixed schedule grew essentially the same harvest on a fraction of the fertilizer. The researchers reached a similar place themselves. They pointed to the recommendation step, not the leaf reading, as the part to fix next, which leaves the camera side of the system as something worth building on and the dosing side as something to correct before an app like this decides how much fertilizer goes on.