Statewide Herbicide Resistance Screening and AI-Based Identification of Horseweed (Erigeron canadensis) in Alabama

Abstract

This thesis evaluated potential herbicide resistance in Alabama horseweed (Erigeron canadensis L.) populations and compared two digital detection approaches for horseweed identification using hyperspectral imagery. The first study focused on evaluating 68 horseweed populations collected across Alabama for resistance to 2,4-D, dicamba, glyphosate, and saflufenacil and documenting their distribution throughout the state. The results indicate that horseweed has developed 31-fold resistance to 2,4-D, 39-fold resistance to dicamba, and 79-fold resistant to glyphosate. Saflufenacil provided complete control for all tested populations. Out of total populations tested, 19% were putatively resistant to 2,4-D, 14% putatively resistant to dicamba and 43% putatively resistant to glyphosate. Glyphosate-resistance was conferred by Pro106Ser substation in EPSPS-2 gene while no such target-site mutation was found in Aux/IAA gene families. Instead, Cytochrome P450, ABC/PDR transporters, and Glutathione S-transferase (GST) are suspected to be candidate genes for NTSR resistance, but confirmation is contingent on further analysis with replications. These results indicate the increasing challenge in managing horseweed using glyphosate and synthetic auxins in the southeastern United States, highlighting the need for diversifying herbicide mode of actions to reduce the risk of resistance evolution and control failure. The second study compared the spectral profile-based weed detection and a deep learning-based segmentation approach for horseweed detection in cotton and peanut under controlled environmental conditions. The results demonstrate that context-aware binary spectral classification improved the classification accuracy to 67.7% compared to 45% in three-class spectral classification model. This reduced efficiency suggests that reflectance signatures alone are insufficient to reliably separate multiple species, likely due to spectral overlap among species. Whereas YOLOv8l-seg model resulted in more reliable discrimination of all the three species- cotton, peanut, and horseweed with mAP50–95 of 0.643, a mask mAP50 of 0.838 and a box mAP50–95 of 0.747, demonstrating model's robust segmentation and object localization performance. Taken together, the results suggest that YOLOv8l-seg is better suited than spectral classification for multi-species discrimination because of its superior robustness and scalability. Spectral classification remains valuable for binary discrimination tasks and may enhance performance when combined with spatial features; however, deep learning-based object detection and segmentation are more appropriate for real-world applications involving diverse and co-occurring weed species.