Explore the Abilities and Potentials of Deep Transfer Learning on High-resolution PolSAR Image Interpretations

Inhalt:
Deep learning (DL) is the hottest technology of the big data era, and the state-of-the-art performances of various nat-ural and optical image interpretation tasks have been achieved by deep convolutional neural networks (DCNNs). To explore the abilities and potentials of DL, we solve a scene classification problem through transfer learning based on different kinds of DCNNs trained with the Imagenet dataset. We would like to answer “What kind of features can a advanced DCNN learn for PolSAR images? How are they different from the ones obtained for photos or optical im-ages? Will specific filter layers learn speckle noise and are the learned features smooth?” during the experiment.