Field-speed plant stress detection from spectral data

Challenge

Challenge

Lab-based plant tissue analysis is accurate but slow and costly. Growers needed stress diagnosis in the field, on the spot, from a handheld sensor.

Approach

Approach

We trained classifiers on smartphone-coupled visible-near-infrared spectra: weighted k-nearest neighbor, discriminant analysis, support vector machines, and ensembles, with principal component analysis for feature extraction. Validation combined controlled greenhouse experiments with independent multi-state field trials.

Deliverables

What the client received

• Multi-class classifier separating healthy, water-stressed, and nutrient-deficient plants
• Deployment-ready model for smartphone-based field use
• Validation study across multiple crop types

Outcome

Outcome

Classification reached 92.3% accuracy with a false positive rate under 8%. Field diagnosis returned in under 30 seconds at 85% lower cost than lab analysis. Presented at Resources for Future Generations 2018, Vancouver.