YOLO11-SCC: Causal-guided semi-supervised deep learning for strawberry maturity detection

• Build causal graph via biological mechanisms to optimize model structure. • Improved C3k2, causal attention and loss for enhanced feature expressiveness. • Semi-supervised training with 30 % labeled data performs well in experiments. • Across diverse test sets, excel in accuracy, robustness, and generalization. Strawberries, a vital crop in global agroeconomics and nutrition, relies on precise preharvest maturity assessment to secure postharvest quality, shelf life, and commercial value. However, traditional assessment methods suffer from subjectivity, inefficiency, and sample damage, failing to meet modern agricultural demands. To address these challenges, this study introduces YOLO11-SCC—a semi-supervised deep learning framework that integrates causal inference—to reduce labeling costs, enhance generalization, and eliminate spurious correlations in strawberry maturity detection. Built using the YOLO11 architecture as a benchmark, YOLO11-SCC constructs a causal graph via biological mechanisms and incorporates it to optimize its structure. It integrates the C3k2_SynTex module to interactively model color and texture features, boosting their synergistic expressiveness. Simultaneously, it embeds a causal attention mechanism into the PSABlock module, building a causal-guided feature optimization pathway to enhance feature discriminability and training stability. Furthermore, it adopts a causal weight matrix-guided CMLoss, allowing it to focus on core causal relationships. Employing a semi-supervised framework, YOLO11-SCC reduces labeling dependency by using only 30 % labeled data. Experiments show 91.8 % precision, 91.2 % recall, 95.3 % mAP50, and 79.7 % mAP50–95 achieved on a greenhouse strawberry dataset. Compared to mainstream models, YOLO11-SCC excels in accuracy and anti-interference performance—especially under complex backgrounds and challenging shooting angles—while achieving an optimal speed-precision balance. Moreover, it exhibits minimal performance fluctuations across historical data, UAV-captured, and ground-planted strawberry test sets, demonstrating exceptional robustness and generalization. This research contributes to advancing the application of deep learning in agriculture.