DSH_YOLO enables factory farming production: a method for detecting and grading the cumulative internode length of tomato seedlings

Introduction: The industrial cultivation of tomato seedlings requires a high degree ofuniformity and consistency in grading. However, traditional grading methods based on phenotypictraits such as leaf area and canopy width are susceptible to environmental conditions, therebylimiting the accuracy and efficiency of grading. Since cumulative internode length is relativelystable and closely correlated with seedling vigor, this study aims to develop an accurate methodfor measuring and grading the cumulative internode length of tomato seedlings. Methods: A tomato seedling cumulative internode length detection method, termedDSH_YOLO (Deformable Convolution-SIoU-Haar Wavelet Downsampling YOLO), wasproposed based on the YOLOv8-seg framework. First, deformable convolution was introducedinto the backbone to enhance feature extraction for curved stems and occluded regions. Second,the original loss function was replaced with SIoU to improve the alignment between predictedregions and the actual stem structure. Third, a Haar wavelet downsampling module was embeddedinto the backbone to preserve high-frequency detail information and reduce information loss underocclusion conditions. An intelligent grading system was further developed to verify the practicalapplicability of the proposed method. Results: Experimental results showed that DSH_YOLO achieved a Precision of 96.1%, aRecall of 94.3%, and an [email protected] of 92.1%. Compared with the baseline YOLOv8 model, theproposed method substantially improved segmentation performance for cumulative internoderegions in tomato seedlings. In prototype validation, the intelligent grading system achieved anaverage grading success rate of 87.50%, with an average cumulative internode length error of 8.0mm. Discussion: The results indicate that DSH_YOLO and the grading system can meet therequirements for large-scale grading and detection of tomato seedlings, demonstrating highdetection accuracy and success rates. This approach can provide insights for grading other types ofseedlings during their growth stages and offer support for seedling production and thedevelopment of intelligent agricultural equipment.