This is the authors' abstract. We don't add key points for this paper.
BACKGROUND: Image recognition tools for weed identification, such as smartphone applications, have the potential to enhance user knowledge, provide early invasive weed alerts, and enable site-specific weed management. Although numerous studies have reported product development and model accuracy, few have evaluated these tools in practical environments beyond developmental settings. In this study, we developed a weed detector using the You Only Look Once (YOLO) v3 object detection algorithm to identify six noxious weed species. Specifically, we examined the effects of: (i) image collection locations, (ii) target backgrounds, and (iii) camera devices on detection accuracy, assessing applicability through field verification at 68 sites across Japan and controlled garden experiments. RESULTS: Detection success was influenced by the background of the target species in images, with significant interaction effects observed between background and target species on detection outcomes. In the most affected combination (background: tray; species: Ipomoea lacunosa), the average precision (AP) value decreased by ~0.2 compared with the other conditions. AP values in field tests were lower than those from test data resembling training data, with no correlation between AP values from test data sets and field verification. No clear effects of land use or camera devices on detection success were detected. CONCLUSION: This study highlights the importance of background in image-based weed detection and identifies limitations in detector applicability. Our findings are expected to support more-efficient development and underscore the importance of sharing applicability data for improved weed detection tools.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Matsuhashi et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: