Abstract: In order to solve some key problems existing in the previous prediction model for cultivated land area, for instance, the nonlinearity, sparseness and the result reliability, a new prediction model for cultivated land area is proposed through fusing and improving the self-adaptive evolution and relevance vector machine in the paper. By analyzing the characteristics of convergence rate for differential evolution algorithm, the functional relationship among shrinkage ratio factor, crossover probability, maximum fitness and minimum fitness is established. Meanwhile, individual shrinkage ratio factor and crossover probability of the next generation are determined based on the current individual fitness data. And the self-adaptive differential evolution algorithm is also developed in this way which can effectively improve the global convergence ability and the robustness of the algorithm. The current studies have confirmed that the kernel parameters have a greater impact on the prediction accuracy of relevance vector machine. Therefore, in order to improve the accuracy of the model, fitness function is established based on leave one cross validation and the relevance vector machine on the basis of self-adaptive differential evolution is also proposed by optimizing the kernel parameters. As the new model has the advantages of sparsity and nonlinearity, and can output the information of the uncertainty of the results, the new method is used to predict the cultivated land area. In order to prove the excellent properties of the new method, the accuracy of the model is evaluated by choosing 5 kinds of precision indices including mean absolute error (MAE), mean relative error (MRE), root mean square error (RMSE), posterior error and error frequency. The computational efficiency and the reliability of the model are estimated quantitatively by running time and confidence interval. Taking Huangshi City as an example, a short-term and a middle-term prediction model for the cultivated land area are set up on the basis of the self-adaptive evolution and relevance vector machine. And these two established prediction models are also compared with the multivariate linear regression model, back propagation (BP) neural network and least squares support vector machine in terms of accuracy, computational efficiency and reliability. The experimental statistics reveal that the newly established prediction model based on the self-adaptive evolution and relevance vector machine is about 2 times higher than the rest 3 models in accuracy, 2 times as much as multivariate linear regression model and 2 orders of magnitude higher than the BP neural network and least squares support vector machine in computational efficiency; the actual land area of the test data set is all in confidence intervals at the 95% confidence level, which is obtained by prediction model for cultivated land area based on self-adaptive differential evolution and relevance vector machine. All the above data confirms that the model based on the self-adaptive evolution and relevance vector machine is a new approach to the prediction of the cultivated farm land with high accuracy, fast calculation and strong reliability.