Search

• Heterogeneous feature learning. Exploiting the complementarity of numerical data and text data to address the issue of data sparsity. • Novel multi-task model. A Shared-bottom layer is introduced between the MoE layer and the input layer of MMoE to reduce feature dimension. • Practical application. The output layer of Shared-MMOE is modified to multi-classification mode, enabling the diagnosis of 17 kinds of diseases and pests and the recommendation of 32 kinds of medicines. • Sequential dependence handling for cascaded task. In order to improve the recommendation effect of medication tasks, the output of diagnostic task is taken as the input of medication task in the form of probability weighting. Research on diagnosis of crop diseases and pests becomes a hot topic of the application of artificial intelligence technology in smart agriculture. Plant electronic medical records (PEMRs) formed by Beijing Plant Clinic provides a new idea for the diagnosis and prevention of crop diseases and pests. PEMRs are stored in the form of heterogeneous data, containing a wealth of plant information, disease and pest information, and environmental information. Therefore, it is urgent to mine the information in PEMRs and employ it to assist in intelligent prescription recommendation. This paper divides prescription recommendation into two sub-tasks, diagnosis and medication, and transforms this problem into a recommendation problem based on multi-task learning, with the goal of establishing a single model to realize learning multi-task simultaneously. Firstly, the correlation analysis of tasks and features is carried out using methods such as knowledge graph. Further, according to the sequential dependency between tasks, a novel cascaded multi-task crop prescription recommendation method based on Shared-Bottom and MMoE (Shared-MMoE) model is proposed, and each task is optimized by gating network. A PEMRs dataset containing 8 diseases, 9 pests and 32 medicines was constructed for model verification. Compared with the baseline model, the experiments showed that Shared-MMoE could significantly improve the quality and accuracy of prescription recommendation. The AUC of diagnosis task and medication task reached 96.33% and 95.36%, respectively. In conclusion, our study preliminarily explored the potential application of artificial intelligence in the research of crop diseases and pests based on PEMRs and multi-task learning. [ABSTRACT FROM AUTHOR]