[Objective] A model will be established to address the issues existing in current soil moisture prediction models， such as insufficient spatial feature mining， limitations in temporal dynamic modeling， and weak adaptability to data distribution differences in the collaborative optimization of complex spatiotemporal dynamic representation and cross-scenario sample generalization. This will provide a scientific decision-making tool for dynamic early warning of drought and flood disasters， precision allocation of irrigation resources， and improved agricultural disaster resistance. [Methods] To address these issues， an attention-guided spatiotemporal feature dynamic fusion network （AGSMP-Net） was proposed. The model integrated a long short term spatiotemporal prediction network with a feature-time-space attention mechanism module， enabling focused processing of time-series information and the capture of spatial distribution variations. It identified the long-term variation trends of soil moisture and optimized the utilization of information across spatiotemporal dimensions. [Results] Experiments validated the feasibility of the AGSMP-Net model in predicting soil moisture using meteorological factors （precipitation and soil temperature）. In the soil moisture prediction task for Henan Province from 2015 to 2024， compared to ConvLSTM， AGSMP-Net model improved the accuracy （R²） from 0.758 to 0.806 and reduced the root mean square error （RMSE） from 0.069 to 0.057. Precipitation has a significant effect on the prediction accuracy of soil moisture model. [Conclusion] The proposed model dynamically allocates feature weights through the spatiotemporal attention mechanism， which can effectively capture both the abrupt responses and steady-state trends in soil moisture variations， thereby improving the accuracy of soil moisture prediction.