Abstract
Precision agriculture relies heavily on effective weed management to ensurerobust crop yields. This study presents RoWeeder, an innovative framework forunsupervised weed mapping that combines crop-row detection with anoise-resilient deep learning model. By leveraging crop-row information tocreate a pseudo-ground truth, our method trains a lightweight deep learningmodel capable of distinguishing between crops and weeds, even in the presenceof noisy data. Evaluated on the WeedMap dataset, RoWeeder achieves an F1 scoreof 75.3, outperforming several baselines. Comprehensive ablation studiesfurther validated the model's performance. By integrating RoWeeder with dronetechnology, farmers can conduct real-time aerial surveys, enabling precise weedmanagement across large fields. The code is available at:\url{https://github.com/pasqualedem/RoWeeder}.