Development of a weed management system for precision farming

The primary objective of this research project is to develop a system for precision spraying of herbicides in a corn field. Ultimately, such a system would permit real-time image collection, processing, weed identification, mapping of weed density and sprayer control using a tractor-mounted digital camera and on-board computer. The initial hypotheses underlying this project were (1) that it is possible to train an artificial neural network (ANN) to distinguish weeds from a crop species (corn in this study); (2) that it is possible to differentiate between weed species; and (3) that precision spraying can significantly reduce the quantity of herbicide needed to protect crop yields, thus reducing both the costs and environmental impacts of such applications. Thus, development of an ANN for this purpose was the main focus of the research project.Since the success of ANN development is primarily dependent on the type of information that it is provided, much of the work involved investigation of different approaches to extracting information from the digital images of field sections and individual objects (weeds or corn plants), as well as analysis of the type of information extracted. The applicability of a given image processing method was evaluated in terms of the image recognition accuracy, as well as the computer time and memory requirements for processing and obtaining ANN output, since speed is of the essence in real-time applications. The greenness method based on a pixel-by-pixel analysis of red-green-blue intensity value of the original images was the most successful and was used in further work.As it turned out, ANN development for this purpose was difficult. While the success rate for recognition of corn plants was high (80% or greater), the success rate for recognition of weeds tended to be low. Improvements in weed recognition were met with decreases in the success rate of corn recognition. Differentiation between weed species was less than desirable. Differentiation between corn and a given weed species was also not as good, particularly when the weed species was similar in appearance to the young corn plant.Therefore, another strategy was developed to recognize weeds in the field by taking images between the corn rows. Previously, the images were taken randomly in the field. The images were processed to obtain percent greenness in each image and this information was used to create weed coverage and weed patchiness maps. Based on these maps, herbicide spraying was decided and spraying amounts were determined. In terms of real-time, it was possible to process the equivalent of one metre of row per second. Although this is slow compared to tractor speed in the field, the computer was not operating under dedicated conditions as one would require for the real-time application. Thus, the results were considered encouraging.The final stage of the work involved an evaluation of the potential herbicide savings from a precision spraying system. This was done by using the weed coverage and weed patchiness maps as inputs to a simulated fuzzy logic controller, and integrating the output of the controller over the field area corresponding to the input images. The simulations with different fuzzy rules and membership functions indicated that the precision spraying approach could reduce the amount of herbicide needed for weed control in a corn field by up to 15%