DDCANet for SiO₂-mediated drought regulation research: high-precision segmentation and phenotypic detection of...

Cucumber is a core cultivated facility vegetable in China. Drought stress at the seedling stage severely inhibits its growth and development. The regulatory mechanism and optimal application concentration of SiO₂ nanoparticles in alleviating drought stress in cucumber seedlings remain unclear. Moreover, traditional manual measurement and classic point cloud segmentation models struggle to achieve high-throughput accurate detection of cucumber seedling phenotypes under drought stress. To address these issues, this study focused on phenotypic detection under drought stress and analysis of the regulatory effects of SiO₂ nanoparticles. An improved compact and low-redundancy segmentation model, DDCANet, was proposed based on PointNet++-SSG. Combined with 3D point cloud technology and the Euclidean clustering algorithm, it enables automatic extraction of phenotypic parameters from cucumber seedlings treated with SiO₂ nanoparticles under drought stress. In this study, Trailing cucumber seedlings were used as experimental materials. 3D point cloud data of cucumber seedlings were collected under treatments with different concentrations of SiO₂ nanoparticles and PEG-simulated drought stress. A dataset containing 70 valid samples was constructed and labeled into two categories: Stem and Leaf. The core optimizations of the DDCANet model are as follows: Firstly, an Adaptive Density-Aware Feature Enhancement (ADFE) module is embedded to accurately capture point cloud density heterogeneity induced by SiO₂;Secondly, a Channel Attention and Normalization-enhanced SA Layer (CANL) is designed to strengthen the coupling of local and global drought-related phenotypic features. Thirdly, a Drought-Aware Hybrid Loss (DHL) function is constructed to alleviate the class imbalance of seedling stem and leaf point clouds under drought stress. Results show that the DDCANet model achieves a mean Intersection over Union (mIoU) of 89.01 ± 0.32% and a Stem IoU of 83.6 ± 0.45%, representing improvements of 6.55% and 9.6% respectively compared with the baseline PointNet++-SSG model, and a 30.5% improvement in stem segmentation accuracy compared with the classic PointNet model. It thus enables high-throughput, non-destructive detection of drought phenotypes in cucumber seedlings under SiO₂ nanoparticle treatment. Ablation experiments verified the positive contributions of the ADFE, CANL, and DHL modules. Furthermore, instance segmentation and phenotype extraction were completed using the Euclidean clustering algorithm to analyze the drought-alleviating effects of SiO₂ nanoparticles under PEG-simulated drought stress. Results indicate that a low concentration of 20 mg/L exhibits a weak alleviating effect, medium concentrations of 40-60 mg/L show bidirectional regulatory characteristics, and a high concentration of 100 mg/L causes negative physiological effects. The optimal application concentration is 80 mg/L, which comprehensively improves key phenotypes such as seedling height and volume under drought stress and exerts a positive regulatory effect on seedling growth under drought conditions. The DDCANet model constructed in this study provides an efficient technical tool for the accurate phenotypic detection of crop seedlings treated with SiO₂ nanoparticles under drought stress. It clarifies the optimal application concentration of SiO₂ nanoparticles, offers a precise concentration threshold and theoretical support for the scientific application of SiO₂ nanoparticles in drought-stressed cultivation of protected cucumber, and establishes a novel methodological reference for the research on phenotypic regulation of crops under drought stress via nano-agricultural technology.