Aiming at the problems of huge agricultural data, low utilization rate, complex structure and fragmented knowledge in China, a top-down and bottom-up agricultural knowledge map construction method was proposed. Focusing on the four elements of crop varieties, crop diseases and insect pests, crop introduction, and model methods, the model layer was constructed from the top down, and the conceptual framework of the knowledge graph was formed through ontology modeling, the data layer was constructed from the bottom up, through data acquisition, knowledge extraction, and fusion, storing and establishing the relationship between entities. Aiming at the problem of ambiguous fields in the corpus, this method collects large number of proprietary vocabularies in the construction of knowledge graphs to segment and mark them. In order to solve the problem of multi-word in agricultural knowledge, many main crop aliases were collected and assigned as entities. Bi-LSTM-CRF was used for named entity recognition, and LSTM was used to classify the problem, and TF-IDF was used for keyword extraction, and finally the knowledge was stored in the Neo4j graph database. The research can be used for agricultural knowledge intelligent retrieval systems, intelligent search systems and other applications to improve user experience.