Multispecies weed mapping using deep learning on UAV imagery for SSWM in maize and tomato

Abstract Accurate identification and mapping of multiple weed species at early growth stages is a critical step toward operational site-specific weed management (SSWM), yet most UAV-based studies have so far been limited to broad weed categories or single dominant species. This study evaluated the potential of deep learning models, including three convolutional neural networks (Inception-ResNet-v2, EfficientNet-B0, YOLOv8) and two Vision Transformers (ViT-Base, Swin-T), to classify, detect and map nine common weed species in maize and tomato fields using UAV-based RGB imagery. The two best-performing classifiers were then implemented in object detection frameworks (YOLOv8m and DETA), and species-specific treatment maps were generated using adaptive economic weed thresholds applied to gridded density weed data. Classification results showed that Swin-T and YOLOv8 achieved the highest metrics, with weighted F1-scores of 98.1% and 97.0%, respectively. Next, the YOLOv8m provided the most accurate and efficient detection, with a mean Average Precision of 0.93 and recall of 0.94, while substantially reducing inference time. The multispecies treatment maps revealed over 70% of weed-free areas, indicating the potential benefits of cost-saving approaches compared to uniform full-field treatments, providing valuable inputs for decision support systems and smart sprayers to gradually advance SSWM for a more selective, efficient and sustainable weed control.