Mushroom growth stage detection and maturity prediction using lightweight vision and logistic models for smart cultivation

Accurate growth stage detection and prediction of full-maturity time are critical for optimising yield, scheduling harvests and improve labour efficiency in mushroom cultivation. However, existing vision-based methods mainly address stage classification without addressing temporal prediction. To overcome this limitation, this study proposes a lightweight detection–prediction framework (YOLO-Pulmonarius, YOLO-P) specifically designed for oyster mushrooms. A Swin Transformer v2 module was integrated into the YOLOv12 detector to improve long-range feature modelling and enhance robustness against glare, occlusion, and overlapping fruiting bodies. Temporal prediction of full maturity was achieved through logistic curve fitting coupled with recursive Gaussian fusion, incorporating area normalisation and dynamic priors for adaptive temporal modelling. The proposed framework was trained on 3,867 annotated images across three growth stages of grey oyster mushrooms. The proposed YOLO-P detection model achieved an average mAP@50 of 97.2%, with the highest performance of 98.6% in the Expired stage. Logistic modelling revealed a stable growth inflection at 62 h (≈2.6 days). The full-maturity time prediction achieved a mean absolute error of 0.42 days and a Pearson correlation of 0.86, outperforming ablation variants without area normalisation (0.49 days, R = 0.79) and without dynamic priors (0.52 days, R = 0.82). The proposed framework establishes a closed-loop pipeline that integrates stage detection with temporal prediction. Its lightweight architecture enables deployment on edge devices, offering a practical solution for real-time monitoring and automated scheduling in smart mushroom cultivation.