Abstract: Existing crop yield prediction methods, often driven by huge historical data for precise estimation, hardly reflect the complexities and uncertainties of crop yield system. Thus, such models could not be used to predict crop yield by using scanty data or to adequately determine phenological events. Assuming that only known yield levels are need for decision-making, Credal network model (which better processes uncertainties) was introduced for predicting crop yield systems. This paper predicted yield levels and determined low yield indicators with scanty data. The paper analyzed the relationship between crop yield and affecting factors and proposed the concept of class element. Also an impact factor concept was proposed to describe the importance of the elements. Furthermore, a general crop yield prediction model based on Credal network was constructed, illustrating the relationships among crop class elements and impact factors. In the constructed prediction Credal network model, every node was assigned with many states treated as probability events. Every state was connected with a conditional or marginal probability representing relationships of variables called parameters. Parameters were denoted as forms of Credal sets to describe system uncertainty. By using forward inference capabilities of Credal network along with “crop yield” variable as objective function, states probabilities of occurrence were calculated with incomplete evidence. Then probabilities of state events higher than others were used in yield forecasts. The process reckoned with knowledge reasoning in the face of scanty facts. By further using backward inference function of Credal network, diagnosis of low-yield facts was accomplished with incomplete evidence. It was noted that the key impact factors and class elements triggered low yield. Combined with a scenario of climate and production as the main factors, interaction was demonstrated with approximation inference algorithm based on extended relational database model. The model illustrated how forecasting and diagnosis based on Credal network could be carried out for crop yield. The inference result showed that the method was feasible, reasonable and interpretable. It fully demonstrated the model functions such as forecasting, interpreting and diagnosis in crop yield systems. The crop yield prediction model based on Credal network enriched existing methodology for crop yield prediction. If in addition to climate and production factors, bio-agronomic, economic measurement and plant growth factors were taken into account, the model provided an all-around method with accurate yield forecasting. The model also supported scientific decision-making in agricultural productions to avoid detrimental factor effects.