SAR image classification with convolutional neural network using modified functions

Identifying a target accurately in the presence of noise in synthetic aperture radar (SAR) images poses significant challenges considering various parameters, such as viewing angle and configuration changes. In this paper, SAR images are classified using the proposed convolutional neural network. The limitation of data concerning moving and stationary target acquisition and recognition (MSTAR) database led us to develop a primary network (learning network) by employing the proposed N-Sigmoid function and training it using eight different satellite databases with images more than MSTAR. We saved the best weights for the main network. Next, we replace the weights of the two fully connected (FC) layers of the ResNet-50 with the saved weights of the training network and employed the main network for classification. Our proposed method’s primary contribution is twofold, based on the availability of additional information in heterogeneous SAR images and differences in intensity and clarity. First, we utilize the weights learned by the training network, and second, we keep negative neurons and neurons close to zero active instead of deactivating them to facilitate better training of the main network. Our proposed method outperformed other references in terms of sensitivity in all classes and achieved an accuracy of 98