.png){kind=icon}
.png){kind=icon}
Return
UAV-Based Multi-Label Weed Detection for Site-Specific Weed Management Using a Multi-Scale Convolutional Attention Network
.png){kind=icon}
OA
AI
M
I
J
DOI：10.3390/agronomy16161544.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate weed detection is essential for site-specific weed management because uncontrolled weeds compete with crops for water, nutrients, light, and growing space, while uniform herbicide application increases production costs and environmental pressure. Existing UAV-based weed detection methods remain limited by three core gaps: inadequate representation of heterogeneous and imbalanced UAV-derived agricultural records, insufficient joint modeling of crop–weed visual similarity and weed-scale variation, and evaluation protocols that rarely assess complete multi-label agreement together with field perturbations and cross-dataset transfer. To address these gaps, this study proposes AgroWeed-MCANet, a multi-scale convolutional attention network for UAV-based multi-label weed detection. Robust feature refinement and compact feature-map arrangement stabilize heterogeneous inputs; convolutional patch encoding, multi-scale ConvNeXt extraction, dilated context aggregation, channel–spatial attention, and cross-scale fusion address local ambiguity, scale variation, and noisy sensing conditions; and the Multi-Label Agreement Score (MLAS), robustness analysis, and cross-dataset evaluation provide field-oriented performance validation. Experiments were conducted on the UAV-UndesirablePlant-UK dataset containing 209,600 labeled field observations with naturally imbalanced crop and weed categories. AgroWeed-MCANet achieved 96.9% accuracy, 95.8% precision, 95.2% recall, 95.5% F1-score, and a 0.912 multi-label agreement score, outperforming twelve recent UAV-based weed detection baselines. Robustness analysis showed stable performance under missing observations, illumination shifts, altitude variations, and 20% noise, with the model maintaining 95.3% accuracy and 93.8% F1-score. Cross-dataset evaluation on DeepWeeds and CWFID further confirmed its transferability, with accuracies of 91.8% and 90.9%, respectively. These findings demonstrate that AgroWeed-MCANet supports reliable UAV-assisted weed monitoring and contributes to sustainable, site-specific weed management in precision agriculture.
Keywords:
precision agriculture
UAV remote sensing
weed detection
site-specific weed management
multi-label classification
convolutional attention network
sustainable agriculture
Journal
A
IF:
3.4
Papers:
1.7W
Citations:
5.0W
Organization
Cited Papers