A Levenberg-Marquardt neural network model with rough set for protecting citrus from frost damage

The protection of citrus from night frosts is a recurrent and important issue that has been researched for many years. Although some feasible methods can be used to protect against and prevent frost, they should be implemented before the frost actually occurs. Therefore, how to accurately predict the temperature change in advance is a core problem for protecting citrus from frost damage. This paper proposes a new method, which combines the neural network with rough set based on the conditional information entropy, in order to improve the accuracy of temperature prediction. Utilizing attribute reduction drawing on the theory of rough set, the weak interdependency in the neural network can be decreased and the prediction accuracy can be increased. Some experiments show that the ability of a neural network to accurately predict minimum temperature can be improved through attribute reduction.