Early-stage recognition of common ragweed (Ambrosia artemisiifolia L.) using low-cost UAV RGB imagery and deep learning

• Optimal common ragweed detection during vegetative stage with U-Net++ IoU of 0.6647. • Low-cost UAV RGB imagery at 0.3 cm resolution enables fine feature extraction. • Weighted loss function mitigates class imbalance for sparse ragweed plants. • Phenology-driven accuracy: high in vegetative, low in flowering, recovery in mature stage. Common ragweed ( Ambrosia artemisiifolia L.) is a globally invasive plant. Its accurate recognition during the vegetative growth stage is crucial for implementing early control measures. However, this task is highly challenging due to the plant's morphological similarity to related species, its sparse and irregular distribution within natural communities. Existing studies are either limited by image resolution or rely on costly hyperspectral sensors, making economically effective large-scale monitoring difficult to achieve. This study reveals that the phenological stage is a core factor governing recognition accuracy. Through systematic analysis of low-cost unmanned aerial vehicle (UAV)-acquired sequential RGB images with a 0.3 cm resolution, we found that recognition performance fluctuates regularly across vegetative, flowering, and mature stages: peak identification occurs during the vegetative stage (U-Net++ IoU: 0.6647), characterized by simple vegetation structure and clear target features, constituting a key time window for control; performance declines significantly during the flowering stage due to severe occlusion from canopy closure; recognition capability partially recovers during the mature stage as reproductive structures provide supplementary features. Addressing the core difficulty of sparse foreground targets in the vegetative stage, we successfully mitigated class imbalance using a weighted loss function, significantly enhancing the model's segmentation capability for sparse ragweed plants. This study confirms that by precisely leveraging phenological timing rather than enhancing sensor configuration, a deep learning approach based on low-cost RGB imagery can fully meet the accuracy requirements for early monitoring, providing a reliable and economical technical paradigm for the efficient control of invasive plants.