A UAV-based tobacco plant detection model integrating NDVI and multi-scale feature fusion for precision agriculture

• Proposed YOLO NTD, the first NDVI-guided UAV detection framework for transplanted tobacco seedlings. • Designed SOEP, FCM, and FPSC modules to improve multi-scale fusion and small-object representation. • Integrated Normalized Wasserstein Distance (NWD) for robust tiny-object localization beyond IoU metrics. • Achieved state-of-the-art accuracy (69.9 % mAP@50–95) with lightweight complexity (1.03 M params, 7.7 GFLOPs). • Demonstrated the potential of vegetation index–deep learning integration for precision agriculture monitoring. Accurate and early detection of tobacco plants is essential for optimizing field management and ensuring stable yield in precision agriculture. Yet, achieving reliable detection at the transplanting stage using Unmanned Aerial Vehicles (UAVs) is particularly challenging due to complex soil backgrounds and the prevalence of small, obscure targets. This study aims to develop a robust UAV-based detection framework that integrates vegetation indices with deep learning to enhance discrimination between crops and non-crops during the critical transplanting stage. We propose YOLO NTD, an innovative detection framework that incorporates NDVI (Normalized Difference Vegetation Index) spectral information into a deep learning pipeline. The architecture integrates three dedicated modules: (i) Small Object Enhanced Pyramid (SOEP) for capturing fine-grained features, (ii) Feature Complementary Mapping (FCM) for enriching multi-scale contextual information, and (iii) Fusion and Pyramid Spatial Channel (FPSC) for optimized feature fusion. Additionally, the Normalized Wasserstein Distance (NWD) metric is introduced to reduce localization sensitivity in small-object detection. Experimental results show that YOLO NTD achieves state-of-the-art performance, reaching 69.9 % mAP@50–95 and 54.6% APtiny, significantly surpassing the baseline model while maintaining low computational overhead. These findings confirm the efficacy of combining vegetation indices with deep learning for enhanced small-object detection. This study provides a reliable and efficient solution for UAV-based crop monitoring, demonstrating that the integration of spectral indices with advanced detection models can substantially improve precision agriculture practices, particularly in early-stage crop management.