Figure 1. People hand hoeing in the fields
Ross D. Lamm, David C. Slaughter, D. Ken Giles, Won Suk Lee,
Biological & Ag. Engineering, UC Davis
Ron Vargas, Agronomy & Range Science, Shafter Research Station & UC Davis
Row-crop farmers have a critical need for rapid, inexpensive and environmentally safe methods of weed control in the seed line. Both hand hoeing and selective herbicides are the traditional methods used for weed control between cotton plants within a row. However, the labor costs associated with hand hoeing, as well as the economic and environmental costs associated with herbicide spraying, have elevated to a nearly prohibitive level. The goal of this project is to utilize machine-vision technology in order to develop a real-time weed control system that can detect crop and weed locations, remove weeds, and thin crop plants during the early stages of growth when weed control is most critical.
Cotton has been the major agronomic field crop in California for years. California alone has traditionally produced over 1 million acres of cotton with a crop value of almost $1 billion (1987 dollars) annually (Kempen 1989).
Effective weed control maximizes cotton yield. The negative effects caused by weeds in cotton farming are two fold: initially, in the early stages of the season, weeds compete with crop plants for light, space, nutrients, and water. Later in the season during harvest, weeds decrease the value of the cotton by depositing trash into the lint or by simply staining the lint, thus causing discoloration (Vargas et al. 1996).
The cost of weed control varies greatly depending on location and weed species, but usually falls between $60/Acre and $150/Acre (1987 dollars) (Kempen 1989). Considering that the total cost of production ranges between $600/Acre and $900/Acre, weed control is a sizeable percentage of the total cost of production in cotton.
Farmers need alternatives for weed control to reduce chemical use and production costs. Recent advances in computer technology and machine vision, present exciting possibilities to reduce or eliminate chemical application through the use of a robotic weed-control system.
One of the most challenging problems for machine vision in agriculture is individual plant identification for weed-control in the seed line. Several research projects have been conducted for machine-vision identification of non-occluded plants in a laboratory environment (e.g. Franz et al. 1991; Guyer et al. 1994). For a robotic weed control system, to be practical, it must have reliable recognition rates of individual plants in a multitude of possible field situations. The actual conditions in the field include touching (occluded) plant and weed leaves, non-uniform soil conditions, changing sunlight throughout the day, insect leaf damage, and wind that, in Northern California, is frequently strong enough to blow plants over. Not only are the actual field conditions complex, the system must work at speeds ranging from 1-3 mph to be commercially acceptable.
Several researchers have used shape features for leaf identification; however, the two main drawbacks of these systems have been occlusion and processing speed. Lee et al. (1998) developed a real-time weed control system for tomatoes, which used color machine vision. The system relied heavily on leaf shape features, and therefore was computationally intensive. Lee et al. report an 81% and 95% recognition rate for tomatoes and weeds respectively at a travel speed of .77mi/hr under ideal, non-occluded conditions. For a real-time weed control system to be acceptable to cotton growers in California, the farmers would need a system that could travel at or above 2mi/hr.
The goal of this project is to utilize machine-vision technology in order to develop a real-time weed control system that can detect crop and weed locations, remove weeds, and thin crop plants during the early stages of growth when weed control is most critical.
Two machine vision algorithms and the illumination chamber were refined in 1998 and implemented on a real-time machine vision system to differentiate cotton seedlings and weeds. These two techniques are well suited for weed control since they can overcome leaf overlap and can be implemented in real-time. The goal of the real-time prototype for 1999 is to analyze 6 images per second, thereby allowing the system to identify and spray weeds at tractor speeds between 1-2 mph.
last modified 12/17/2003