Soil fertility is the comprehensive reflection of related factors and the related factors. Soil fertility evaluation knowledge is stored by relational database as usually, and it is difficult to show the correlation and constraints among attributes .In this paper, Nongan county farmland productivity data is as the research object, Using rough set approach to do attribute reduction, using ontology method to establish the soil fertility level knowledge base, using multi Agent technology to implement the prototype system, and complete the reuse and sharing of knowledge, lay the foundation for semantic level reasoning.