Method for estimating single-leaf photosynthesis in crops considering few-shot learning and uncertainty prediction

Accurate and efficient assessment of single-leaf photosynthetic rate (A_L) is essential for applications in crop yield assessment, stress perception, and biological breeding. However, existing A_L prediction methods typically require extensive training data and often rely on point estimation, which poses challenges for few-shot learning and high-reliability predictions. To alleviate this dilemma, a comprehensive A_L estimation model was developed in this study. Using the combination of meta-learning (ML) and multilayer perceptron (MLP), the model initially constructed a base prediction model and then fine-tuned it for unseen tasks. Compared to the current popular methods, the proposed model achieved higher prediction accuracy with fewer training samples. Introducing only about 5 % of new samples from the dataset was sufficient to achieve satisfactory predictive performance. Additionally, this study introduced quantile regression (QR) method to obtain the 95 % confidence interval of A_L, mitigating the limitations of high reliability in predictions. Finally, by employing Gaussian kernel density estimation (GKDE) to derive the probability density under each environmental condition, we constructed an A_L prediction model with reliable uncertainty estimation capabilities. Through detailed validation using multiple datasets from various species and interpretability analysis, the proposed method has proven to be universally applicable and reasonable. This research highlights the high generalization ability of ML-MLP for unseen datasets and extends A_L predictions from point estimation to interval prediction by QR-GKDE, thereby facilitating the researches on crop cultivation. This research significantly enhances precision agriculture by providing robust methodologies for crop monitoring and stress perception.