In order to accurately achieve the effective ground information in the process of remote sensing image processing and shorten the classification time, the convolutional neural networks (CNN) model is introduced into the classification of remote sensing image features. First, the picture fuzzy weighted average (PFWG) improved CNN classification method is proposed. The fuzzy geometric clustering algorithm is used as a pre-processing unit to characterize the experimental samples, and for multi-source feature decision-making for remote sensing ground information. The classification process is simplified and the convergence of the CNN model is speeded up. The experimental results show that using PFWG improved CNN classification method, the overall classification accuracy reaches 93.73%, and the Kappa coefficient is 0.94. This method effectively compensates for the shortcoming of CNN itself which is not good enough for classification and has poor expression performance of remote sensing images. It has successfully completed an efficient classification task and has a certain anti-jamming capability.