An experiment was conducted to determine the utility of multispectral imagery for identifying soybean, bare soil, and six weed species commonly found in Mississippi. Weed species evaluated were hemp sesbania, palmleaf morningglory, pitted morningglory, prickly sida, sicklepod, and smallflower morningglory. Multispectral imagery was analyzed using supervised classification techniques based upon 2-class, 3-class, and 8-class systems. The 2-class system was designed to differentiate bare soil and vegetation. The 3-class system was used to differentiate bare soil, soybean, and weed species. Finally, the 8-class system was designed to differentiate bare soil, soybean, and all weed species independently. Soybean classification accuracies classified as vegetation for the 2-class system were greater than 95%, and bare soil classification accuracies were greater than 90%. In the 3-class system, soybean classification accuracies were 70% or greater. Classification of soybean decreased slightly in the 3-class system when compared to the 2-class system because of the 3-class system separating soybean plots from in the weed plots, which was not done in the 2-class system. Weed classification accuracies increased as weed density or weeks after emergence (WAE) increased. The greatest weed classification accuracies were obtained once weed species were allowed to grow for 10 wk. Palmleaf morningglory and pitted morningglory classification accuracies were greater than 90% for 10 WAE using the 3-class system. Palmleaf morningglory and pitted morningglory at the highest densities of 6 plants/m(2) produced the highest classification accuracies for the 8-class system once allowed to grow for 10 wk. All other weed species generally produced classification accuracies less than 50%, regardless of planting density. Thus, multispectral imagery has the potential for weed detection, especially when being used in a management system when individual weed species differentiation is not essential, as in the 2-class or 3-class system. However, weed detection was not obtained until 8 to 10 WAE, which is unacceptable in production agriculture. Therefore, more refined imagery acquisition with higher spatial and/or spectral resolution and more sophisticated analyses need to be further explored for this technology to be used early-season when it would be most valuable.
展开▼
