A mechanistic structure-based hybrid CNN-LSTM model for greenhouse tomato yield prediction

: Greenhouse crop yield prediction is challenging due to the complex influence of multiple environmental factors to crop growth. To effectively integrate mechanistic knowledge and advanced modeling methods, and simultaneously enhance the accuracy, adaptability, and dynamic performance of greenhouse yield models, this paper proposes a mechanistic structure-based hybrid Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) model, in which a Mechanistic Structure (MS)-based hierarchical data pattern is proposed by decomposing the common structure of mechanistic models, and CNN and LSTM modules are constructed to integrate the learning of the effect of multidimensional greenhouse climate factors on yield data and the dynamic relationship between yield data. Then, the yields calculated from a mechanistic model are hybridized with the yield values calculated from the MS-CNN-LSTM. Finally, by mapping the extracted features to the target values and updating all trainable parameters by backpropagation, the yield model of the target greenhouse is learned adaptively. Experiments and analysis on data originating from two real tomato greenhouse systems, (a) multi-year records from a same greenhouse and (b) same time but with different climate control and management strategies, validated the accuracy and stability of the proposed model, and a set of in-season prediction experiments are designed to further validate the dynamic performance of this model. This proposed method provides a more accurate, stable and dynamic yield modeling approach for intelligent greenhouse systems compared to current yield models.