Yield prediction models can be improved by better integration of data and algorithms, and the accuracy of yield prediction can be further improved by incorporating other factors such as those affecting yield into the model. The research situation was addressed that the wheat-maize rotation system lacked the direct incorporation of the previous crop information into the yield prediction and management of the seasonal crop, a multi-temporal and multimodal crop yield prediction model based on GPR was established by using remote sensing information of the growing season and yield information of the previous maize crop, fusing multi-temporal and multimodal data such as remote sensing information of wheat growing season at the jointing stage, filling stage and maturity stage, fertilization information before sowing and soil properties. The results showed that the performance of the yield prediction model based on the multiple growth periods was improved compared with that based on the single growth period, in which the decision coefficient R2 of the yield prediction model was improved by 0.01~0.03. The accuracy of the yield prediction model based on the spectral indexes of wheat growing season at the jointing stage was higher than the accuracy of the yield prediction model based on the spectral indexes of wheat growing season at the filling stage, and the accuracy of the yield prediction model based on the spectral indexes of wheat growing season at the maturity stage was the lowest, and the accuracy of the yield prediction model based on the jointing stage was slightly lower than that of the yield prediction model based on the multiple growth periods, but the accuracy was similar. In the yield prediction models based on the multimodal parameters fusion, the yield prediction models based on two-modal parameters fusion had higher accuracy than the unimodal yield prediction models, except for the yield prediction model constructed by fusing maize information with soil properties. The accuracy of the yield prediction models with four-modal parameters fusion and three-modal parameters fusion was higher than that of the corresponding yield prediction models with low-modal parameters fusion. The GPR model with four-modal parameters fusion had a decision coefficient R2 of 0.92 and RMSE of 213.75kg/hm2, which improved R2 by 0.02 to 0.41 compared with the wheat yield prediction models based on other modalities. For wheat yield prediction models based on multimodal parameters fusion, from large to small, the influence of each modal parameters was as follows: fertilization information, wheat remote sensing information, soil properties information, maize crop information. Maize crop information had the least improvement in the accuracy of the yield prediction models based on the multimodal parameters fusion, which improved R2 by 0.02~0.07. Maize crop information characterized the soil fertility condition of post-harvest to a certain extent, and it was a high spatial resolution supplement to soil properties information, which could further improve the ability to quantify soil fertility, then combined with other parameters, they can improve the accuracy of wheat yield prediction. In conclusion, the research result provided a scientific basis and method for the comprehensive utilization of soil-crop data and the comprehensive management of wheat-maize rotation system.