Real-time, machine-learning weed detection system for autonomous agricultural machines

点击翻译按钮获取中文摘要

Current agricultural practices include blanket application of herbicides for eliminating unwanted plant speciesbefore they can steal resources and cause harm to agricultural cash crops cultivated by farmers. Applyingherbicides to entire fields instead of individual plants is wasteful. However, identifying and targeting individualweeds has not been feasible in large-scale food production. While much progress has been made, thesetechniques are still not robust enough to accurately differentiate between weeds and crops, especially inreal-world conditions consisting of multiple weed species within a single field. Thus, the aim of this EngageGrant is to create a real-time system to classify digital images of weeds and crops, where the goal is reducingthe amount of herbicide used in food production, ultimately leading to increased crop yields and lower costs.The focus of the proposed system is to classify the crop species of corn and canola as well as the weeds Canadathistle, cocklebur, redroot pigweed, and dandelion using deep neural networks. The benefits includeenvironmental footprint reduction for Canadian farmers, and lower food production costs.