What we’ve tested.
What we’ve learned.

Our instrument · Idaho, July 2026

A six-band sensor, tested in the field.

We built a handheld multispectral sensor and tested it on 99 plants at two Idaho sites. The initial pooled analysis found a signal associated with visual PVY symptom labels.

Plants

99

Spectral bands

6

Pooled ROC-AUC, approximately

0.78

What it tells us

A purpose-built instrument can capture a detectable field signal. This first trial also identified measurement problems that matter more than adding more samples to the same setup.

Limitations

Labels were visual symptom calls, not laboratory-confirmed infections. Brightness and sensor position dominated the measurements. Calibration issues included clipped reference readings and physically implausible reflectance in two bands. These data do not establish detection before visible symptoms.

The sensor findings are ours. The later RGB photo analysis is a separate experiment, summarized below.

Public dataset · Wageningen, 2017

Plant-level signal in tractor imagery.

We analyzed the hyperspectral dataset associated with Polder and colleagues’ 2019 paper. Our initial analysis kept repeated visits to the same plant together when evaluating the model.

Plant-visits in the initial cohort

3,346

Independent PVY plants

68

Grouped ROC-AUC

0.886

What it tells us

In this dataset, five or six carefully placed bands performed similarly to the full spectrum. Coarser spatial resolution caused only a modest reduction in ranking performance. These findings inform the next instrument design.

Limitations

Ranking plants is different from reliably identifying infected plants. The initial analysis had a precision–recall area of 0.195 at about 3.9% prevalence. Even the highest observed precision on its curve was 24%, at 68% recall. Those results still produce substantial false alarms.

Band placement is a hypothesis to test on our instrument. It has not been independently validated across the other datasets, and the fitted models did not transfer between datasets.

The measurements belong to the original researchers. The numbers above describe our analysis, not the performance reported by the paper’s authors.

Original study: Polder et al., 2019

Public dataset · Montana State

The evaluation split changes the answer.

We analyzed the UAV hyperspectral dataset associated with Nesar and colleagues’ 2025 paper. The trial inoculated plants by plot, making plot identity and disease label difficult to separate.

Random-split ROC-AUC

0.827

Whole plots held out: ROC-AUC

0.596

Group-level permutation p-value

0.11

What it tells us

Good performance on a random split can overstate how well a model will work in an unseen plot. The plot-held-out result was not statistically significant. It does not establish operational UAV detection.

Limitations

The delivered cubes contain 223 bands from approximately 407–899 nm, with a gap around 747–770 nm. Plot confounding and limited independent groups constrain the conclusions. More patches from the same plots are not equivalent to more independent field trials.

This is our analysis of public data. It is not a reproduction of the authors’ reported performance.

Original study: Nesar et al., 2025

Retrospective follow-up · September 15, 2026

More information helps. Validation still matters.

A follow-up study examined V1 field photographs and richer plant-level spectral summaries. Evaluation that accounts for representation and model selection produced ROC-AUC of 0.891 for V1 RGB and 0.898 for the reconstructed Netherlands cohort.

The V1 photos contain the symptoms used to assign the labels. That result supports recognizing visible symptoms; it does not establish laboratory-confirmed or presymptomatic PVY detection. The reconstructed Netherlands cohort also differs slightly from the initial cohort above, so the two sets of numbers are reported separately.

In the Netherlands follow-up, a fixed distribution-and-local-context model achieved 17.3% precision at 90.2% recall. That is still too many false alarms for an automatic removal decision.

Read the follow-up report

What comes next

The next instrument needs consistent calibration, controlled measurement geometry, and plant-level laboratory labels. It must be evaluated on the intended platform, in new fields, against other diseases and stresses.

Late blight and stress categories on the landing-page dashboard are interface concepts. The farm map and totals are fictional. They are not evidence of validated detection or completed field work.

Reading the metrics: ROC-AUC measures how well a model ranks positives above negatives; it is not the percentage of plants correctly diagnosed. Precision is the proportion of flagged plants that are positive. Recall is the proportion of positive plants found. Field prevalence changes the precision a grower should expect.