Pscilly

Mushroom Crop Stress AI: Penn State Model Spots Damage in Button Mushrooms

By · Published

A new artificial intelligence tool for spotting mushroom crop stress was built and tested on Pennsylvania's signature crop. Researchers at Penn State have developed a model that reads hyperspectral images of white button mushrooms and flags damage with high accuracy, while using far less computing power than a comparable AI system. The team published the work in the journal Scientific Reports.

The project grew out of a practical problem for growers and packers. Mushrooms bruise, heat up and break down quickly after harvest, and much of that deterioration starts before anyone can see it. Catching it early could mean less waste and better quality on store shelves.

Why white button mushrooms

The researchers chose Agaricus, the familiar white button mushroom, for a reason close to home. According to the U.S. Department of Agriculture, Pennsylvania grows 74% of the nation's white button mushrooms, and national mushroom sales dropped 16% to $922 million last season. The USDA has also identified postharvest quality loss as a significant challenge for the industry.

That matters to the farms that anchor the state's mushroom economy, from Phillips Mushroom Farms in Kennett Square to specialty growers such as Mycopolitan Mushrooms in Philadelphia and Bounty Beneath Mushrooms in Worthington. Any tool that helps sort good product from damaged product sooner could eventually reach packing lines across the region.

How the model works

Hyperspectral cameras capture light across many narrow wavelengths, revealing chemical and structural changes in plant tissue that the human eye misses. The trouble is volume: tracking those images over time produces enormous datasets that are slow and expensive to process.

Senior author Shirin Ghatrehsamani, an assistant professor of agricultural and biological engineering in Penn State's College of Agricultural Sciences, said the team responded by designing a model it calls PSUMamba. It splits the work into three tracks. One reads spectral signals tied to biochemical changes, a second follows how conditions shift over time, and a third looks at spatial patterns within each image. The results of all three are then combined.

First author Weiqun Wang, a doctoral candidate in the same department, said crops often show subtle physiological changes well before visible symptoms appear. The goal, he explained, was a model accurate enough to catch those signals but light enough to run in real agricultural settings where decisions need to happen fast.

Putting mushrooms under stress

In a greenhouse facility at the university, the team deliberately damaged batches of white button mushrooms to mimic common postharvest problems. Some were dipped in 140-degree Fahrenheit water for 30 seconds to imitate heat exposure during processing. Others were microwaved for 30 seconds to cause localized cell damage. A third group was punctured, scraped and knocked around to replicate rough handling. A control group was kept under the same conditions without treatment.

The researchers then captured 204 bands of light from the mushrooms at eight points in time and asked the model to tell the groups apart.

The results

According to the team, PSUMamba reached an average accuracy of 94.47%, a sensitivity of 98.98% and an AUC-ROC score of 99.36%, a standard measure of how well a classification model separates categories. It also needed roughly 10 times fewer computational operations than a Vision Transformer, the image-analysis model used as a benchmark.

Tests that switched off individual tracks showed that the spectral, time-based and spatial views each contributed something different, and that the combination performed best.

What comes next

The study was run under tightly controlled lab conditions, and the researchers say more work is needed before the approach is proven in the field. Ghatrehsamani said the next step is to test the model on real farms, across more crops and types of stress, and to see whether it can be paired with greenhouse monitoring platforms, field sensors, robotic systems and possibly aerial imaging.

For Pennsylvania's mushroom industry, the takeaway is an early but promising one: a sensing approach designed around the state's biggest specialty crop that could one day help growers catch quality problems before they reach the consumer.

PSUMamba performance detecting mushroom stress
Accuracy94.47 %
Sensitivity98.98 %
AUC-ROC99.36 %

Penn State researchers, Scientific Reports (2026); lab-controlled tests on white button mushrooms

On Pscilly

Published
November 22, 2026
Place
State College, PA

Nearby