Use of artificial intelligence to locate and treat weeds in Midwestern United States corn (Zea mays) and soybean (Glycine max) cropping systems

Abstract

Site-specific weed management (SSWM) is defined as the process of managing weeds where they are growing as opposed to treating the whole field and treating areas with no weeds. Artificial intelligence (AI), the process of creating intelligent machines, has become a part of everyday life in modern society. Utilizing convolutional neural networks and object detection algorithms, weeds can be distinguished from crops, and herbicide applications can target weeds where they are growing. The objectives of this dissertation were to 1) train open-sourced object detection algorithms to detect in central Kansas soybean (Glycine Max [L.] Merr.) fields, focusing on Palmer amaranth (Amaranthus palmeri S. Watson, henceforth denoted as A. palmeri), 2) determine herbicide efficacy and cost savings of SSWM herbicide applications using a ONE SMART SPRAY research sprayer, an intelligent dual-boom sprayer using AI technology to locate and spray weeds growing within crops, and 3) compare traditional broadcast (BCST) applications with spot-spray (SS) herbicide applications using a commercial-sized ONE SMART SPRAY sprayer. Images were obtained from two soybean fields in 2021 containing A. palmeri infestations and were annotated with bounding boxes to identify both A. palmeri and soybean plants. In this study, the YOLOv5 object detection algorithm was identified as having the highest mean average precision scores and was therefore selected for further analysis. The precision, recall, and F1 evaluation metrics found for the test image dataset was 0.71, 0.70, and 0.71, respectively. Regression analysis revealed that our trained YOLOv5 model evaluation metrics were higher when identifying A. palmeri plants 2 cm in height at low plants m-2. For the second objective, corn (Zea mays L) and soybean field trials were conducted in Manhattan, KS and Seymour, IL with the research-sized ONE SMART SPRAY. Simultaneous herbicide applications of residual BCST + foliar SS, base-rate foliar BCST + SS “Spike” rates, and SS only were compared in corn (Zea mays L.) and soybean trials. Specific SS thresholds tested included an herbicide efficacy, balanced, savings, and traditional BCST applications were tested for comparison. Results showed that both residual BCST + foliar SS and “Spike” approaches provided weed-free area not different than traditional broadcast applications, in many cases. The greatest savings were achieved by SS only applications, but weed-free area was almost always significantly less than for other treatments. Simultaneous BCST + SS of soil residual and foliar herbicides, respectively, provided the most weed-free area with the greatest cost savings for both crops. Thirdly, we tested a commercial-sized ONE SMART SPRAY and compared traditional broadcast applications with SS only and simulated two-boom/two-tank applications using the foliar base rate + “Spike” approach. Treatments included SS only, low rate BCST + high rate SS, and high rate BCST + low rate SS applications. Results indicated that high rate BCST + low rate SS applications provided the highest weed-free area, but savings were not different from a broadcast application in soybeans. In corn, thresholds were not different, but both applications with BCST applications demonstrated greater weed-free area than SS only treatments. Overall, this research demonstrated that 1) open-sourced object detection algorithms can be custom trained to identify A. palmeri in soybean crops, with opportunities to train and identify other weed species in other crops, 2) intelligent AI sprayers show potential in providing weed-free area comparable to traditional BCST applications, especially systems that utilize two-tank/two-boom technology for simultaneous BCST and SS applications, and 3) herbicide costs were significantly reduced for SS applications compared to traditional BCST applications.