Skip to main content
ARS Home » Research » Publications at this Location » Publication #120151
Title: DIGITAL IMAGING FROM AGRICULTURAL AIRCRAFT - RESOLUTION AND WEED CLASSIFICATION ISSUES USING ENVI 3.2

Author
item Thomson, Steven
item Bryson, Charles

Submitted to: Proceedings of Southern Weed Science Society
Publication Type: Abstract Only
Publication Acceptance Date: 2/23/2001
Publication Date: 2/26/2001
Citation: THOMSON, S.J., BRYSON, C.T. DIGITAL IMAGING FROM AGRICULTURAL AIRCRAFT - RESOLUTION AND WEED CLASSIFICATION ISSUES USING ENVI 3.2. PROCEEDINGS OF SOUTHERN WEED SCIENCE SOCIETY. 2001. 54:210-211.

Interpretive Summary:

Technical Abstract: Several classification techniques available in the ENVI 3.2 image analysis software were compared to classify images obtained from a Sony digital video camera, mounted in a spray plane. Preliminary attempts were made to classify weeds, crop, and soil. Four comparisons between methods were made, along with a comparison of four classification methods on early cotton. The green band of a non-filtered image was analyzed. The two supervised classification methods used (Parallelepiped and Mahalanobis Distance) successfully separated Johnsongrass, hyssop spurge, and cotton in an image taken early in the season. The Mahalanobis Distance method with properly programmed parameters detected more hyssop spurge, consistent with field scouting. In cases where Johnsongrass was highly visible, bidirectional effects accounted for most of the system's ability to distinguish between Johnsongrass and crop. The Isodata (unsupervised) method was unable to distinguish between weeds and crop when bidirectional effects were not pronounced, as did simple thresholding in many cases. The digital video system worked well for this preliminary study, although image resolution could be improved to discern smaller weed patches early in the season. An alternative might be to fly at lower altitudes for greater resolution. We found, however, that separation of classes suffered when flying low, as areas approached greater homogeneity. This was probably because the camera averages light intensity from the entire scene to set the aperture. Using the digital video camera, classification yielded better results for the unfiltered images (green band) than pre-filtered near-infrared (NIR) images. When using ENVI, simple thresholding seemed to be appropriate for analyzing gray-scale NIR images.