Real-time UAV-based wheat lodging detection via edge-accelerated improved Mask-RT-DETR

To address the challenges of slow response, low efficiency in manual field inspection, and frequent missed detection of lodging areas in traditional monitoring of wheat lodging disasters, this study proposes a drone-based wheat lodging inspection method leveraging an improved Mask-RT-DETR model and edge computing. Considering the complex characteristics of wheat lodging in aerial images—such as diverse stem bending angles, severe occlusion within plant populations, and complicated background interference—three key improvements are introduced to the Mask-RT-DETR model. First, a bottleneck convolutional kernel optimization module is designed, employing 1 × 1 channel compression and 3 × 3 depthwise separable convolutions to enhance the extraction of spectral features at stem fracture points. Second, a Cascaded Group-wise Attention (CGA) module is embedded into the Transformer decoder to strengthen the spatial correlation modeling of stem inclination angles and canopy density. By combining multi-head attention mechanisms with a cascaded group strategy, CGA reduces computational load while improving feature representation. Third, the localization loss function is upgraded to Focal-EIoU Loss (Focal Efficient Intersection over Union Loss), which, together with a dynamic sample matching strategy, enhances regression accuracy of lodging bounding boxes and reduces model parameters while preserving multi-scale feature fusion capability. A dataset is constructed using field imagery collected by a drone platform equipped with high-resolution RGB and multispectral sensors. Experimental results show that, compared to the original model, the improved model achieves a 6.7 percentage point increase in precision and reaches a detection speed of 63.2 FPS. In cross-model comparative experiments, the proposed method outperforms Faster R-CNN, SSD, YOLO series models, and the original Mask-RT-DETR in precision (P), recall (R), mean average precision (mAP), and F1-score, achieving a lodging detection accuracy of 97.2%. To evaluate edge computing performance, the model is deployed on a Jetson Orin Nano embedded device. After acceleration with TensorRT, it achieves 96.3% accuracy, 96.5% recall, and a real-time detection speed of 32.0 FPS, satisfying the requirements for real-time, high-accuracy field monitoring. The results demonstrate that the improved model maintains a lightweight architecture while significantly enhancing detection accuracy, providing an effective technical solution for drone-based wheat lodging inspection that balances detection performance and computational efficiency.