Yinglei Cheng

Recently, convolutional neural networks (CNNs) have shown significant advantages in the tasks of image classification; however, these usually require a large number of labeled samples for training. In practice, it is difficult and costly to obtain sufficient labeled samples of polarimetric synthetic aperture radar (PolSAR) images. To address this problem, we propose a novel semi-supervised classification method for PolSAR images in this paper, using the co-training of CNN and a support vector machine (SVM). In our co-training method, an eight-layer CNN with residual network (ResNet) architecture is designed as the primary classifier, and an SVM is used as the auxiliary classifier. In particular, the SVM is used to enhance the performance of our algorithm in the case of limited labeled samples. In our method, more and more pseudo-labeled samples are iteratively yielded for training through a two-stage co-training of CNN and SVM, which gradually improves the performance of the two classifiers. The trained CNN is employed as the final classifier due to its strong classification capability with enough samples. We carried out experiments on two C-band airborne PolSAR images acquired by the AIRSAR systems and an L-band spaceborne PolSAR image acquired by the GaoFen-3 system. The experimental results demonstrate that the proposed method can effectively integrate the complementary advantages of SVM and CNN, providing overall classification accuracy of more than 97%, 96% and 93% with limited labeled samples (10 samples per class) for the above three images, respectively, which is superior to the state-of-the-art semi-supervised methods for PolSAR image classification.

Distance measure plays a critical role in various applications of polarimetric synthetic aperture radar (PolSAR) image data. In recent decades, plenty of distance measures have been developed for PolSAR image data from different perspectives, which, however, have not been well analyzed and summarized. In order to make better use of these distance measures in algorithm design, this paper provides a systematic survey of them and analyzes their relations in detail. We divide these distance measures into five main categories (i.e., the norm distances, geodesic distances, maximum likelihood (ML) distances, generalized likelihood ratio test (GLRT) distances, stochastics distances) and two other categories (i.e., the inter-patch distances and those based on metric learning). Furthermore, we analyze the relations between different distance measures and visualize them with graphs to make them clearer. Moreover, some properties of the main distance measures are discussed, and some advice for choosing distances in algorithm design is also provided. This survey can serve as a reference for researchers in PolSAR image processing, analysis, and related fields.

The point cloud classification network only uses point cloud structure features or edge features to construct the feature vector, so that the classification accuracy is low. To solve this problem, this paper proposes a multi-level classification network. First, the original point cloud is divided into small samples using the preprocessing algorithm to obtain batch input and improve training efficiency; then construct structural features and edge features for point cloud feature extraction, and obtain local fine-grained descriptions through multi-level expressions within and between points; then design a convolutional neural network for feature learning. With the deepening of the network level, the degree of feature abstraction level is higher and higher, and the degree of distinction increases, so as to effectively improve the accuracy. Using the Vaihingen dataset to test the MLMS-Net network, the accuracy rate reached 85.4%.

Aiming at the classification problem of ground objects in complex scenes of airborne LiDAR data, this paper proposes an algorithm based on three-dimensional convolutional neural network (3D-CNN). The algorithm improves the pre-processing method of LiDAR point cloud and realize automatically classification of LiDAR point cloud in complex scenes. Based on the input of 3D-CNN is voxel grid, this paper selects multi-scales to construct voxel grids for each point in the point cloud, trains them with the network, and then combines the features of multi - scales through the fully connected layer. Finally, the network returns the category score of each point to complete the classification of LiDAR point cloud. The algorithm is verified by the Vaihingen dataset provided by ISPRS. The experimental results show that the proposed algorithm can achieve higher classification accuracy than other convolutional neural networks dealing with point clouds.