Study on neural networks and transfer learning for synthetic aperture radar images change detection

Abstract

In this paper, convolutional neural network-based models for detecting changes in synthetic aperture radar(SAR) images are dealt with. A big model is divided into two models which are for removing speckle, a characteristic of SAR images, from an image and change detection from images. After that, the two models are transfer-learned and selected as a model for detecting changes in the SAR images. The model for removing speckle is selected by the most accurate method among basic deep CNN with different number of parameters and loss functions, and the model for detecting changes is selected from image semantic segmentation models. The speckle removal model selects one of the models with 8 or 16 layers, and the loss function is a combination of mean absolute error(MAE) or mean squared error(MSE) and multi-scale structural similarity(MS-SSIM) error. In the change detection model, the model was selected by comparing the fully convolutional network(FCN), the U-Net++, and the proposed shallow U-Net++, and the distance of dissimilar features was learned through similarity learning using the distance loss function. The dataset for supervised learning of the three models (speckle removal, change detection, and SAR images change detection) appearing in this paper was used by transforming the distributed satellite image dataset. The SAR images change detection model tried to increase the accuracy and speed by using transfer learning using the parameters of the learned speckle removal and change detection model as initial values. When the end-to-end learning model learned by simply connecting the two models and the transfer learning model were trained with the same epoch and time, the model using transfer learning showed better quantitative, qualitative accuracy and learning speed. Even when the model was trained until convergence, the model using transfer learning showed better performance.