Soil temperature is a very important variable in agricultural science, and its spatial and temporal variations are characterized by stochasticity, nonlinearity and non-stationarity, which greatly affects the accuracy of prediction, for this reason, an improved black kite algorithm (IBKA) was proposed to optimize a short-term soil temperature prediction model by integrating the attention mechanism of convolutional neural networks (ACNN) and dendrite networks (DD). Firstly, the black-winged kite algorithm was improved by adaptive hierarchical learning strategy to enhance the optimization ability of the algorithm;then the two models of convolutional neural networks (ACNN) and dendrite networks (DD) with integrated attention mechanism were fused to obtain a new model, ACNN-DD, which was used to mine the relationship between soil temperature and feature variables, and then to output the prediction of the soil temperature in the next 6 hours. Finally, in order to validate the model, soil temperature data monitored at the vegetable planting base in Fengyan Village, Fengyan Township, Nanchuan District, Chongqing, the National Field Scientific Observatory for Naiman Farmland Ecosystems, Inner Mongolia, and the National Field Scientific Observatory for Ansai Farmland Ecosystems, Shaanxi, were brought into the model. The results showed that the coefficients of determination of the model were as high as 0.98, 0.98, and 0.99, and the root mean square errors were as low as 1.12℃, 1.35℃, and 1.37℃, the mean absolute percentage errors were reduced to 3.83%, 5.54%, and 5.41%, which were all superior to that of the traditional soil temperature prediction models, such as ILSTM_Soil, MLP-FFA, and SPA-GA-SVR, etc. This showed that the model can effectively predict the soil temperature in the next 6 hours, which can provide a theoretical basis for the application in the field of smart agriculture.