High-throughput phenotyping of chinese cabbage using multispectral drone imagery and deep learning for morphological, color, and nutritional traits across growth stages

• Cost-effective phenotyping using multimodal data and deep learning models. • Accurate centroid detection at seedling stage with improved YOLOv8 model. • Enhanced multispectral image resolution via SRGAN with transformer attention. • High-throughput quantification of leaf color and morphology traits. • Precise segmentation of multispectral point clouds with improved RandLA-Net. With the increasing demand for precision agriculture, efficient phenotypic analysis is crucial for crop breeding and productivity enhancement. This study presents an economically efficient and high-throughput phenotypic analysis framework for Chinese cabbage ( Brassica rapa L. subsp. pekinensis ), combining low-cost multispectral imaging drones with deep learning technologies. During the seedling stage, we achieved a mAP of 97.0 % and an F1-score of 93.3 %, representing a 12.5 % improvement over baseline models, enabling precise localization of individual plants. At the rosette stage, we employed a multispectral super-resolution generative adversarial network (MSRGAN) to enhance image quality, achieving a canopy segmentation accuracy of 97.99 %, with improvements of 0.83 % in mIoU and 0.63 % in FWIoU compared to baseline models. From the segmentation results, we extracted 23 key phenotypic parameters (e.g., NDVI, EVI, RGB, RE, NIR), which facilitated quantification of leaf color (range: 0–100). These parameters provided support for the prediction of SPAD (R 2 = 0.64) and N content (R 2 = 0.59) across the entire growth period. During the heading stage, we addressed the limitations of 2D imaging for complex 3D structures by achieving 85.01 % mIoU and 92.06 % accuracy in point cloud segmentation, a 10.3 % improvement over existing approaches. Combined with an optimized clustering analysis algorithm, we achieved precise segmentation of individual plants, extracting 33 morphological parameters (e.g., length, width, height) and quantitatively assessing head expansion degree (range: 0–100). The framework demonstrates that this integrated approach, as a practical alternative to traditional field-based methods, could improve the accuracy and efficiency of phenotypic trait extraction for crop monitoring and breeding.