Date of Graduation

12-2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Geosciences (PhD)

Degree Level

Graduate

Department

Geosciences

Advisor/Mentor

Tullis, Jason A.

Committee Member

Spurlock, Terry N.

Second Committee Member

Ham, Richard G.

Third Committee Member

Daniels, Michael B.

Keywords

Geographic information systems; Machine learning; Pesticide application; Precision agriculture; Remote sensing; Unmanned aircraft systems

Abstract

Weed populations in agricultural production fields are often scattered and unevenly distributed; however, herbicides are broadcast across fields evenly. Although effective, in the case of post-emergent herbicides, exceedingly more pesticides are used than necessary. A novel weed detection and control workflow was evaluated targeting Palmer amaranth in soybean (Glycine max) fields. High spatial resolution (0.4 cm) unmanned aircraft system (UAS) image streams were collected, annotated, and used to train 16 object detection convolutional neural networks (CNNs; RetinaNet, Faster R-CNN, Single Shot Detector, and YOLO v3) each trained on imagery with 0.4, 0.6, 0.8, and 1.2 cm spatial resolutions. Models were evaluated on imagery from four production fields containing approximately 7,800 weeds. The highest performing model was Faster R-CNN trained on 0.4 cm imagery (precision = 0.86, recall = 0.98, and F1-score = 0.91). A site-specific workflow leveraging the highest performing trained CNN models was evaluated in replicated field trials. Weed control (%) was compared between a broadcast treatment and the proposed site-specific workflow which was applied using a pulse-width modulated (PWM) sprayer. Results indicate no statistical (p < .05) difference in weed control measured one (M = 96.22%, SD = 3.90 and M = 90.10%, SD = 9.96), two (M = 95.15%, SD = 5.34 and M = 89.64%, SD = 8.58), and three weeks (M = 88.55, SD = 11.07 and M = 81.78%, SD = 13.05) after application between broadcast and site-specific treatments, respectively. Furthermore, there was a significant (p < 0.05) 48% mean reduction in applied area (m2) between broadcast and site-specific treatments across both years. Equivalent post application efficacy can be achieved with significant reductions in herbicides if weeds are targeted through site-specific applications. Site-specific weed maps can be generated and executed using accessible technologies like UAS, open-source CNNs, and PWM sprayers.

Download

Share

COinS