MGS-SR-YOLO: A Cotton Plant Height Measurement Method for Unstructured Cotton-Field Environments

Accurate acquisition of cotton plant height provides a reference for large-scale growth assessment and is important for analyzing local growth variability and abnormal growth. However, in cotton breeding and precision field management, plant height acquisition is inefficient, manual measurement is prone to error, and existing models contain many parameters, limiting deployment on field embedded devices. Moreover, cotton leaf occlusion and complex background interference may cause the loss of key target features, making surveying staffs and scale marks difficult to recognize and segment. To address these issues, this study proposes MGS-SR-YOLO, a lightweight surveying-staff segmentation algorithm for continuous monitoring at local fixed points. MobileNetV4-Small is adopted as the backbone to reduce model parameters. GSConv lightweight convolution and the Sim-VoV-GSCSP module are introduced into the neck network to reduce model complexity while improving target segmentation under occlusion. A learnable sub-pixel convolution-based upsampling module, SRUp, is further constructed to replace static upsampling and preserve target edge details. Finally, HSV color statistical filtering is combined with a geometric difference method to extract the canopy baseline and estimate cotton plant height. Compared with YOLOv11n-seg, MGS-SR-YOLO reduces parameters, computational cost, and model size by 38.0%, 48.4%, and 38.5%, while improving Box mAP50 and Mask mAP50 by 4.4% and 10.2%. In field measurements, the MAE and RMSE are 5.35 cm and 5.50 cm, with RMSE reduced by 15.4% compared with the baseline. The proposed method reduces computational load while supporting automatic plant height estimation in complex unstructured environments.