DIGITAL.CSIC. Repositorio Institucional del CSIC
Publicación
Otros (Other)
Drone-based weed mapping at the species level for precision weed control in maize and tomato fields
Papers presented at the 15th European Conference on Precision Agriculture, Barcelona, Spain, 29 June–3 July 2025, organised by Universitat Politècnica de Catalunya, BarcelonaTech, Accurate weed identification is crucial for effective implementation of precision weed control
strategies. This research evaluated several convolutional neural networks and vision transformer
models aiming to map nine common weed species in drone-based RGB images collected in earlyseason maize and tomato fields. The Droneweed dataset (>65 000 labels) was used to train, validate
and test the models. Swin-T showed most robust performance (98% F1-score) and EfficientNet-B0
the lowest computational cost, though YOLO-v8 achieved superior classification results (97.5%
overall accuracy) and better speed-accuracy balance. Consequently, YOLO-v8m detector was applied
to generate treatment maps, which reported high potential to reduce herbicide usage by 37% and
43% in both fields, respectively., This research was funded by the PID2023-150108OB-C31 and PID2023-150108OB-C32 projects
(MCIU/AEI/10.13039/501100011033, FEDER, EU), and the MMT24-ICA-01 project (MOMENTUM
CSIC program, MCIU/AEI/10.13039/501100011033, PRTR, CSIC). The authors thank the Tech4Agro
technicians’ team (https://tech4agro.csic.es) for their invaluable contributions to field sampling and
weed dataset creation., Peer reviewed