Reflectance data were subjected to a variety of analysis methods to determine the utility of hyperspectral reflectance for differentiating soybean, soil, and six weed species commonly found in Mississippi agricultural fields. Weed species evaluated were hemp sesbania, palmleaf morningglory, pitted morningglory, prickly sida, sicklepod, and smallflower morningglory. Hyperspectral reflectance data were collected from mature plant leaves three times in 2002 and two times in 2003. Vegetation indices were calculated and subjected to principal component analysis (PCA) and linear discriminant analysis (LDA). The PCA, using vegetation indices, produced the poorest classification accuracies for the plant species studied, generally less than 50%, whereas LDA resulted in classification accuracies greater than those from PCA. Best spectral band combination (BSBC) provided the greatest classification accuracies, with all better than 80% for all data sets. The BSBC indicated three wavelength bands of interest for species discrimination in the short wavelength infrared portion of the electromagnetic spectrum, which are not commonly used in current vegetation indices for species differentiation. These areas of interest were located from 1,445 to 1,475 nm, 2,030 to 2,090 nm, and 2,115 to 2,135 nm. The top 10 wavelengths determined by BSBC were then added to the vegetation indices and reanalyzed using PCA and LDA. Classification accuracies increased for all species when these wavelengths were added rather than using vegetation indices alone, suggesting greater crop and weed species differentiation can be obtained when using sensors that include these wavelength regions of the short wavelength infrared portion of the electromagnetic spectrum.

Two experiments, one focusing on preemergence (PRE) herbicides and the other on postemergence (POST) herbicides, were conducted and repeated in time to examine the utility of hyperspectral remote sensing data for discriminating common cocklebur, hemp sesbania, pitted morningglory, sicklepod, and soybean after PRE and POST herbicide application. Discriminant models were created from combinations of multiple indices. The model created from the second experimental run's data set and validated on the first experimental run's data provided an average of 97% correct classification of soybean and an overall average classification accuracy of 65% for all species. These data suggest that these models are relatively robust and could potentially be used across a wide range of herbicide applications in field scenarios. From the data set pooled across time and experiment types, a single discriminant model was created with multiple indices that discriminated soybean from weeds 88%, on average, regardless of herbicide, rate, or species. Signature amplitudes, an additional classification technique, produced variable results with respect to discriminating soybean from weeds after herbicide application and discriminating between controls and plants to which herbicides were applied; thus, this was not an adequate classification technique.