It is widely agreed that a complete Synthetic Aperture Radar (SAR) target feature library should be built to improve the basic support capability of SAR image applications in China.At presentthe accuracy of the SAR image in the SAR target feature library built by the electromagnetic modeling simulation depends on the simulation parameters of the ground object.The simulation parameters can hardly be obtained by theory.To solve this problema method based on Convolutional Neural Network (CNN) is proposed to predict the best simulation parameters of SAR images.An 11-layer CNN regression system is builtwhose input is the SAR simulation image.Since the predicted simulation parameters are 4 dimensional a new loss function is proposed to improve the predicted accuracy of each dimension in the process of multidimensional regression.Through an analysis of the changes in the error amplitude of the parameters during the training of the neural networkit can be seen that the loss function can achieve the desired result in the prediction of all the 4 dimensions.A comparison between the real image and the simulated image shows a high similarity between themwhich validates the effectiveness of this method.