Multispectral CNN-Transformer Framework for UAV-based Weed Detection and Precision Herbicide Recommendation

One of the tasks of precision agriculture is weed detection and selective herbicide application. Multispectral imaging is a high-resolution field monitoring technique provided by UAV, but weed-crop discrimination is difficult due to illumination variation, growth stage difference and fine-grained weed species similarity. WeedNet-X, a hybrid CNN-Transformer framework, is proposed for multi-spectral weed detection, weed species classification (into 6 classes), regression of UGV spray coordinates, uncertainty-aware prediction, federated multi-farm learning, and agronomic decision support using LLM. To enhance weed localization and weed classification, the model adopts multispectral feature encoding, RGB–NIR/red-edge cross-modal attention, and crop-weed attention. Based on the experiments conducted on WeedMap and four custom field datasets with 5-fold cross-validation, WeedNet-X got 94.8% mAP@0.5, 93.2% weed recall, 97.4% crop precision and 96.1% federated accuracy. The system resulted in a 38.6% reduction in herbicide use when compared to blanket spray and an expected calibration error of 0.016. Results show that WeedNet-X can contribute to accurate and actionable weed management in precision agriculture while maintaining privacy.