Toward autonomous weed management systems in sugarcane crops: Are we Close?

Authors: PAPA, JOAO - Sao Paulo State University (UNESP); MANESCO, JOAO R - Sao Paulo State University (UNESP); SCHODER, MICHAEL - Massachusetts Institute Of Technology; JACOBUCCI, CODY - Massachusetts Institute Of Technology; HE, JIANGPENG - Massachusetts Institute Of Technology; Spaunhorst, Douglas; Johnson, Richard; KREBS, HERMANO - Massachusetts Institute Of Technology

Submitted to: Artificial Intelligence
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/16/2026
Publication Date: 3/30/2026
Citation: Papa, J.P., Manesco, J., Schoder, M., Jacobucci, C., He, J., Spaunhorst, D.J., Johnson, R.M., Krebs, H.I. 2026. Toward autonomous weed management systems in sugarcane crops: Are we Close?. Artificial Intelligence. https://doi.org/10.1038/s44387-026-00096-0.
DOI: https://doi.org/10.1038/s44387-026-00096-0

Interpretive Summary: Weeds compete with crop plants for space, nutrients, sunlight, and soil moisture, reducing crop yields, particularly during the first weeks after emergence. Controlling weeds in perennial crops, such as sugarcane, is a challenging undertaking and is typically addressed by the application of preemergence and postemergence herbicides and mechanical tillage. This work focuses on the detection of weeds in sugarcane. We provide an in-the-field dataset as a new benchmark for weed detection and evaluate several computer modeling methods in the three-step process including weed detection, weed classification, and weed separation. Our method achieved a moderate score for its detection ability, a classification accuracy of approximately 99% and evaluated two promising approaches for weed separation. Although considerable progress was made, and the results for all tasks look promising, precisely detecting weeds in perennial crops is not yet a solved problem under real world conditions.

Technical Abstract: Weeds compete with crop plants for space, nutrients, sunlight, and soil moisture, reducing crop yields, particularly during the first weeks after emergence. Controlling weeds in perennial crops, such as sugarcane, is a very challenging undertaking and is typically addressed by the application of preemergence and postemergence herbicides and mechanical tillage. This work focuses on the detection of weeds in sugarcane. We provide an in-the-field dataset as a new benchmark for weed detection and evaluate several deep learning architectures for three downstream tasks including object detection, classification, and segmentation. Specifically, for detection, a 44.2 AP50 score was achieved by combining RTMDeT, an architecture that employs large-kernel depth-wise convolution, a loss function that incorporates geometric constraints, and a detection head composed of feature pyramid networks. For classification, we leveraged Swin Transformer with self-supervised pre-training that achieved around 99% of classification accuracy. Finally, we qualitatively compared the segmentation performance between SAM and ExGR-based approaches. Although significant progress was made and the results for all tasks look promising, precisely detecting weeds in perennial crops is not yet a solved problem in real world conditions.