MPE-YOLO: an efficient and lightweight model for maize plug-tray empty-cell detection and iOS deployment

Empty-cell recognition in maize plug trays is an important component of seedling quality monitoring in large-scale greenhouse nursery production, providing quantitative information for emergence assessment, plug-tray quality evaluation, and automated transplanting operations. Manual inspection is time-consuming and subjective, and automatic recognition remains challenging because greenhouse maize plug trays often contain neighboring-leaf occlusion and substrate-covered cell edges, while practical use also requires lightweight models suitable for mobile deployment.

To address these challenges, this paper proposes MPE-YOLO, a lightweight YOLO26n-based detection model designed for greenhouse maize plug-tray environments. Depthwise separable convolution was incorporated into the backbone to reduce computational redundancy, while CGSB was introduced into the neck to reduce the number of model parameters and computational cost while maintaining feature representation. An EMA-enhanced C3k2 module, termed C3k2-EMA, was used to strengthen attention to empty-cell regions and suppress background interference. In addition, a dual-focusing loss function integrating Focal Loss and Wise-IoU v3 was designed to improve learning from hard-to-classify samples and targets with ambiguous boundaries. The optimized model was further deployed in an iPhone-based mobile application for portable empty-cell monitoring.

Experimental results from repeated tests showed that, on average, the proposed MPE-YOLO outperformed the original YOLO26n, improving precision and mAP@0.5 by 1.6 and 0.3 percentage points, respectively, while increasing inference speed by 16.0%. Stratified evaluation under occlusion and substrate-covering conditions showed that MPE-YOLO maintained better robustness in complex greenhouse plug-tray backgrounds. In empty-cell-free scenes, MPE-YOLO reduced the average number of false detections per tray from 3.1 to 1.3. The iPhone-based application also demonstrated the feasibility of using MPE-YOLO for portable detection.

This work provides an efficient and portable solution for maize plug-tray empty-cell recognition.