An estimated 13% of crops are lost globally to plant diseases. Disease detection, identification, and tracking is performed today by crop scouts, a process that is expensive, slow and difficult, and is impractical to expand to cover all crops. The project involves developing AI-drones that work side-by-side with farmers and identify specific diseases and assess their progress. The use of intelligent drones for crop monitoring will allow farmers to respond quickly to emergent diseases, nutrient stress, and other potentially devastating damages without the prohibitive expense of hiring a crop scout. This ability could increase productivity, and may also help predict, track and respond to epidemics for national and global food security. In addition, this technology will also be used for ongoing collection of precise plant performance data for breeding resistance.
The central hypothesis of this proposal is that drones equipped with trained Convolutional Neural Networks can provide a transformative increase in actionable crop disease identification. A secondary hypothesis is that the proposed phenotyping at the individual plant level will also provide unprecedented resolution of data for future modeling, breeding, and data-driven yield optimization. In order to test this hypothesis, we will develop a UAS platform to collect images over university owned experimental crops, and aim to develop AI to identify pathologies at an accuracy that is on par with human experts. The UASs will consult human experts in ambiguous cases and gradually learn to make decisions autonomously. The key challenge will be development of AI that can reliably diagnose disease with a good accuracy of detection to false alarms. Automatic identification of disease is a challenging machine vision task given the complexity of images, exacerbated by variable lighting and weather conditions, and navigation/stability control.
