Nonlinear Manifold Learning Integrated with Fully Convolutional Networks for PolSAR Image Classification

Abstract

Synthetic Aperture Rradar (SAR) provides rich ground information for remote sensing survey and can be used all time and in all weather conditions. Polarimetric SAR (PolSAR) can further reveal surface scattering difference and improve radar’s application ability. Most existing classification methods for PolSAR imagery are based on manual features, such methods with fixed pattern has poor data adaptability and low feature utilization, if directly input to the classifier. Therefore, combining PolSAR data characteristics and deep network with auto-feature learning ability forms a new breakthrough direction. In fact, feature learning of deep network is to realize function approximation from data to label, through multi-layer accumulation, but finite layers limit the network’s mapping ability. According to manifold hypothesis, high-dimensional data exists in potential low-dimensional manifold and different types of data locates in different manifolds. Manifold learning can model core variables of the target, and separate different data’s manifold as much as possible, so as to complete data classification better. Therefore, taking manifold hypothesis as a starting point, nonlinear manifold learning integrated with fully convolutional networks for PolSAR image classification method is proposed in this paper. Firstly, high-dimensional polarized features are extracted based on scattering matrix and coherence matrix of original PolSAR data, whose compact representation is mined by manifold learning. Meanwhile, drawing on transfer learning, pre-trained Fully Convolutional Networks (FCN) model is utilized to learn deep spatial features of PolSAR imagery. Considering complementary advantages, weighted strategy is adopted to embed manifold representation into deep spatial features, which are input into support vector machine (SVM) classifier for final classification. A series of experiments on three PolSAR datasets have verified effectiveness and superiority of the proposed classification algorithm.

Deep Learning Meets SAR: Concepts, models, pitfalls, and perspectives

“…Chen et al [79] improved CNN performances by incorporating expert knowledge of target scattering mechanism interpretation and polarimetric feature mining. In a more recent work [80], He et al proposed the combination of features learned from nonlinear manifold embedding and applying a fully convolutional network (FCN) on input PolSAR images; the final classification was carried out in an ensemble approach by SVM. In [81], the authors focused on the computational efficiency of deep learning methods, proposing the use of lightweight 3D CNNs.…”

Semi-Supervised Classification for PolSAR Data With Multi-Scale Evolving Weighted Graph Convolutional Network

“…Liu et al [21] proposed a polarimetric convolutional network, where an encoding mode is designed to maintain the polarimetric information of scattering matrix completely. He et al [22] combined nonlinear manifold embedding and FCN, where the learned features are finally fed into SVM to realize ensemble learning and classification. Zhao et al [23] proposed to use two similar dilated FCN frameworks in parallel with different convolutional kernels to extract more discriminate features of PolSAR images.…”

Crop Mapping Using Sentinel Full-Year Dual-Polarized SAR Data and a CPU-Optimized Convolutional Neural Network With Two Sampling Strategies

“…4) The usage of 1×1 convolution: The usage of 1×1 convolutional layer reduces the number of parameters of a deep network, and is useful in deep learning-based remote sensing image classification by easing overfitting [56]. In the pre-deep learning era, remote sensing image classification consisted of two major procedures [57], [58], which often included designing some filter to extracting spatial features (e.g., GLCM) and the fusion of multi-source data (e.g., spectral, spatial, temporal, and backscattering) with dimensional reduction methods (e.g., principal component analysis and manifold learning [59]). If we treat the network into the same two components, then the 1×1 convolutional layer is responsible for the fusion process, and a larger convolutional layer at the head of a network is responsible for spatial feature extraction.…”