Project code: PRR03-550
Project Lead
Anne Smith - Agriculture and Agri-Food Canada
Objective
To evaluate a prototype ground-based camera for mapping of various weed species in canola and field pea plots for use in site-specific herbicide application
Summary of Results
Traditionally herbicides are applied to an agricultural field as a single uniform application. Site-specific herbicide application techniques, which optimize the placement of herbicide, can reduce both the cost of crop production and potential detrimental effects to the environment. To implement site-specific application, information on the location and population density of weed species within a field is required. Over the past 25 years, there have been a number of studies investigating the potential to discriminate crops and weeds using plant spectral characteristics with mixed results. There has been limited exploration into the use of ground based hyperspectral imagery for species discrimination, which may provide improved detection capability due to its high spectral and spatial resolution.
Through funding from Pesticide Risk Reduction, a study was undertaken to examine the use of hyperspectral remote sensing for weed/crop discrimination. This proof-of-concept study was developed with a view to enabling industry to produce herbicide prescription maps for site-specific applications. The study was conducted using three crops (canola, field peas and wheat) and two weeds (redroot pigweed and wild oats) of economic importance in Canada.
In March 2003, a hyperspectral camera was acquired and a ground-based vehicle retrofitted to house the system. Imagery was acquired over a series of small field plots established in 2004 and 2005. A number of problems were identified with the camera system during the 2004 growing season which necessitated interaction with the camera developer and troubleshooting with the manufacturer. New calibration procedures were developed along with both image acquisition and post acquisition processing software programs. These activities were not anticipated and thus were not outlined in the original project proposal. The data from the field season of 2004 and some greenhouse trials conducted in 2005 were used to develop these new calibration procedures.
Methods to discriminate weed and crop species were investigated using the 2005 field plot data collected throughout the season. Neural networks (NN) and maximum likelihood classification (MLC) techniques proved effective in discriminating redroot pigweed in canola, field peas and wheat (overall accuracy ranged from 89 to 96%) on two dates. Classification using the neural networks (NN) approach tended to be slightly more effective than the MLC. Discrimination of wild oats proved more difficult than redroot pigweed, particularly with respect to wild oats in wheat plots as both plants are narrow leaved grass species.
The deliverables, outlined in the original proposal were not fulfilled in the time frame anticipated. However, methods were developed for weed/crop discrimination and a proof-of-concept for weed/crop mapping delivered. An economic analysis of the system potential showed that the benefits of such a system are dependent upon the herbicide costs and the accuracy of the system. The classification results from this study have and will be further shared with the camera manufacturer. However, before this technology can be utilized in the production of weed fraction maps and in the design of herbicide sprayer, further developmental work is required.