• Isik Ozturk

Image - based recognition of weeds and plant stress symptoms

Identification of noxious weeds and phenotyping biotic stresses in plant-pathogen interactions is limited by phenotypes that are not reliably discriminated by visual assessment. Traditionally, weeds and plant disease severity are scored with visual inspection by trained workers. While this approach has been refined over many years and many crop systems, it still is plagued by unreliability. Studies shown that there exists a substantial variation both between individual raters and between different assessments by a single rater.

A variety of quantitative high throughput image-based methods for weed identification and phenotyping plant stress are currently being developed/tested. These methods range from detailed analysis of a single plant over time to broad assessment of the crop canopy for thousands of plants in a field, and employ a wide range of imaging technologies. For weed detection; machine vision technology that employs deep convolutional neural network model was recently found to be exceptionally successful in detecting Poa annua in turf grass. For plant - pathogen interactions, imaging detect spatial patterns of heterogeneity and allows for visualization of localized responses. A computation method based on the combination of multiple images of chlorophyll fluorescence parameters are currently used to characterize and detect plant diseases, except the diseases that progress asymptomatically or with only internal symptoms.

Expert-based approaches of traditional phenotyping are moving toward technology-based approaches providing automatic and quantitative measurements of biotic or abiotic stresses. We, in Herbitech, believe that measurements based on automatic image analysis provide higher accuracy, practicality, and reproducibility than human visual inspections. We are indeed experimenting with various visualization techniques and currently comparing data that are collected by trained worker and digital imaging.

#multispectral #weeddetection #phenotyping