Weed classification is a crucial step in site-specific weed management system that could lead to saving herbicides by preventing repeated chemical applications.

Project personnel planted kochia, buckwheat, green foxtail and Canada thistle weeds in a greenhouse and collected spectral reflectance of those weeds by handheld radio spectrometer to develop spectral signatures. RGB and multispectral images of outdoor weed research plots were collected by UAV to develop an algorithm for classification of weed species.

Project personnel focused on water hemp, kochia and lambsquarter. Reflectance spectra were measured from water hemp, kochia, and lambsquarter plants by a USB2000 Vis/NIR spectrometer when they were less than 3 inches tall. Spectral reflectance of weeds was analyzed by Soft Independent Modeling of Class Analogy. The test set validation results of the SIMCA analysis performed to classify three weed clusters using all spectral information (400-2500 nm) and five pre-processing methods included: 1) multiplicative scatter correction; 2) standard normal variate; 3) vector normalization; 4) first derivative; and 5) second derivative. The discrimination power of different wavelength variables in the model indicated that red and red-edge region (640, 676, and 730 nm) had the best wavelengths for weed discrimination in visible range. Pre-processing of data and SIMCA analysis were all performed with Unscrambler V10.4 software package. Results indicated a multispectral sensor may have potential to identify weed species in fields.  

High resolution RGB and multispectral images over corn and soybean crops were collected using a Phantom3 rotocopter with a Sentera quad sensor that has 6 bands (3RGB, 670,710.730 nm) to identify weeds from crops and classify weed species. Project personnel collected images using a SLANTRANG multispectral camera mounted on Matrice 100 UAV over fields, and prepared mosaics with Agisoft PhotoScan. Before weed classification, row crops were identified by row detection algorithm in Matlab 2016a software. To classify weed species we used maximum likelihood method in supervised classification. This is a pixel by pixel classification method with respect to segmented area. Project personnel compared the result which achieved by image processing method to ground collected data. Weeds were classified successfully more than 90 percentage ground data accuracy. The error was higher when the weeds was less than 3 inches high.

Researchers are continuing to analyze imagery using different supervised and unsupervised classification methods in ENVI to find the most accurate weed species classification method in commercial fields. They are comparing maximum likelihood; decision tree and K- means methods to develop confusion matrix for finding classification error. The plant spectral canopy reflectance extracted from the multispectral images indicated that the spectral signature of weed species is significantly different which can be used to classify weeds with high accuracy.

The research results show clear difference in spectral signature of three weed species including water hemp, lambsquarter and kochia. The Sentera quad sensor multispectral camera at selected wavelength can be helpful to identify weed species in the field.