In recent years, increasing events of  internet public opinion on food security have attracted great attention of the Chinese government. The current early warning index system of internet public opinion on food security lacks a comprehensive consideration of theme attributes, propagation and diffusion indexes, and does not consider the inherent characteristics and evolution law of the public opinion in depth. Moreover, as the current early warning model of internet public opinion fails to take the interrelationship between different characteristics of the public opinion into consideration, which leads to the low accuracy of early warning of public opinion. Aiming at the above problems, we construct an index system consisting of five dimensions, including theme attributes and propagation and diffusion indexes. Based on this, we propose a regularization long short term memory (Re-LSTM) model, which uses the regularization method to constrain the input weight of each unit in the network, and replaces the tanh activation function by the softsign activation function. Compared with other classic models, the proposed model can not only improve the accuracy of early warning, but also better avoid the problem of gradient disappearance and overfitting.