Intelligent detection of wilted cucumber leaves with deep learning model in real greenhouse conditions

To address the issue of leaf water stress and wilting caused by intermittent irrigation in greenhouse cucumber cultivation, which leads to reduced photosynthesis, lower cucumber quality and yield, this study proposes a method for recognizing wilting cucumber leaves based on an improved YOLOv8 deep learning model. This study proposes three innovative improvements to address the three major challenges of leaf recognition in complex greenhouse conditions: 1) high-noise complex backgrounds, 2) feature space mismatch, and 3) multi-scale occlusion problems: a novel Leaf Edge Information Enhancement Module (EIEM) is introduced into the backbone network to enhance the learning of leaf edges and reduce the interference of background noise. Based on this, Contextual and Spatial Feature Calibration Network (CSFCN) is introduced to correct the feature mismatch, and Occlusion-Aware Attention (SEAM) is employed to improve the detection head. Compared to the YOLOv8 model, the YOLOv8-EIEM-CSFCN-SEAM（YOLOv8-ECS）model achieves improvements of 1.8 %, 3.2 %, and 1.8 % in accuracy, recall rate, and [email protected], respectively. Comparison of example tests shows that the average recognition accuracy of the improved YOLOv8-ECS model reaches 97.4 % under different wilting degrees, different angles and different light conditions, which can achieve high-precision recognition of wilted leaves in greenhouse cucumbers under complex backgrounds and occlusion conditions, and provides feasible technical support for realizing intelligent, real-time, and on-demand precision irrigation of greenhouse cucumbers. • A wilted cucumber leaf recognition algorithm based on improved YOLOv8 is proposed. • YOLOv8-ECS model tackles complex backgrounds, feature mismatch and occlusion. • Validates superior recognition performance through comparative analysis.