Measurement and prediction of facility tomato ripening period based on YOLOv11-Seg and LSTM-MHA

In the precise management of facility agriculture, the determination and prediction of tomato maturity stages hold significant importance for optimizing production planning and enhancing economic benefits. However, due to complex growth environments and the dynamic developmental characteristics of tomato fruits, efficiently extracting fruit features and predicting future maturity stages has become a core challenge. To address this, this paper for the first time proposes a fully integrated detection-prediction framework, achieving simultaneous acquisition of current maturity and prediction of future maturity stages during a single inspection. Firstly, an occlusion-optimized segmentation mechanism based on YOLOv11-Seg is developed, which eliminates stem and leaf obstruction interferences through logical operations, enabling complete extraction of tomato peel masks in complex scenes. Subsequently, current maturity is calculated through colour gamut analysis. Building upon this, a dynamic LSTM-MHA prediction architecture is designed, which for the first time combines a multi-head attention mechanism with adaptive sequence processing to accurately capture spatiotemporal characteristics of maturity evolution. Experimental results show that this algorithm achieves significant optimization in tomato fruit detection under stem and leaf occlusion scenarios and obtains high-precision results in future maturity prediction (R² reaches 97.87 %, RMSE value is 3.36 %). Compared to traditional methods, this algorithm markedly enhances both the accuracy of tomato maturity estimation and applicability in complex environments, demonstrating the effectiveness of the YOLOv11-Seg and dynamic LSTM-MHA models. It provides an integrated “detection-analysis-prediction” inspection solution for facility agriculture, promoting agricultural production into a new stage of predictable management.