Using information technology to realize the effective integration and application of multi-source heterogeneous spatio-temporal multi-modal big data of agriculture is a key issue that needs to be urgently solved in the precision agriculture. A method of constructing， controlling and decision-making for precision fertilization was proposed based on the spatio-temporal multi-modal knowledge graph of agriculture to construct a management system and realize the fine management of nutrients in field. The nodes and relationships in the plots to be queried and the spatio-temporal multi-modal knowledge graph of agriculture were embedded and represented through the subgraph matching method based on deep learning. Vector similarity calculation was used to obtain candidate subgraphs. The fertilization model data suitable for query plots were obtained from the information of subgraphs storing historical data. The results showed sub maps isomorphic to the given land query map were obtained in the spatio-temporal multi-modal knowledge graph of agriculture based on the instantiated query map of the land to be fertilized. An agricultural fertilization model suitable for the current plot was obtained from the information of subgraph storing historical decision. It is indicated that the automatic selection of model based on spatio-temporal multi-modal knowledge graph of agriculture is accurate and reliable. It will provide decision-making support for precision fertilization.