Automated plant identification has been studied for decades, with identifying weed species in a crop for targeted weed control as one of the recent major motivations. The two primary sources of information for this task have been spectral data obtained using spectroscopic methods and spatial information from images. These two approaches have only very rarely been incorporated into a single character for identification. An objective of this research was to explore approaches to combining these methods.

An imaging spectrophotometer was designed and built in order to collect the spectral- spatial data. The imaging spectrophotometer combined a spectrograph-based imager with a 1.8 m uniform light source. Illumination sources matched to the input spectrum required to collect data between 400 and 1000 nm. Spectral resolution was approximately five nanometers, with spatial resolution adequate to resolve veins on some leaves.

Data were collected from thirteen crop and weed species at five maturity levels. From these data, characters were developed from the descriptive statistics of the reflectance and reflectance red edge distribution across vegetated areas, spectrally-segmented plant structures, and leaf shape based on idealized templates.

The usefulness of these characters for classification varied. Spectrally-segmented plant structures (fleshy tissues and red stems) were very good at reducing the list of potential candidates, reliably indicating those samples exhibiting the character from those that did not. Characters based on reflectance distribution statistics (reflectance skewness at 896, 511, and 713 nm, and standard deviation at 730 and 959 nm) and red edge location and slope statistics (red edge slope skewness, slope standard deviation, and slope mean, and red edge location standard deviation and location mean) demonstrated some ability to distinguish between species, but require further development of their definition and meaning. Leaf segmentation was done using an edge subtraction approach using a 5.2 nm waveband centred at 719 nm, achieving a mean leaf segmentation rate of 63%. Leaf shape characters performed poorly in testing, primarily a result of bias in methods of combining evidence for shape. While performance was mixed, all of the prototype characters illustrate possible directions for the development of characters incorporating both spectral and spatial information.