Search

Support vector machine (SVM) is a powerful cognitive and learning algorithm in the domain of pattern recognition and image classification. However, the generalization ability of SVM is limited when processing classification of high-resolution remote sensing images. One chief reason for this is that the Euclidean distance-based distance matrix in traditional SVM treats different samples equally and overlooks the global distribution of samples. To construct a more effective SVM-based classification method, this paper proposes a multi-scale Mahalanobis kernel-based SVM classifier. In this new method, we first introduce a Mahalanobis distance kernel to improve the global cognitive learning ability of SVM. Then, the Mahalanobis distance kernel is embedded to the multi-scale kernel learning (MSKL) to construct a novel multi-scale Mahalanobis kernel, in which the parameters are optimized by a bio-inspired algorithm, named differential evolution. Finally, the new method is extended to the classification of high-resolution remote sensing images based on the spatial-spectral features. The comparison experiments of five public UCI datasets and two high-resolution remote sensing images verify that the Mahalanobis distance-based method can obtain more accurate classification results than that of the Euclidean distance-based method. In addition, the proposed method produced the best classification results in all the experiments. The global cognitive learning ability of Mahalanobis distance-based method is stronger than that of the Euclidean distance-based method. In addition, this study indicates that the optimized MSKL are potential for the interpretation and understanding of complicated high-resolution remote sensing scene.

Hyperspectral image classification has been widely used in a variety of applications such as land cover analysis, mining, change detection and disaster evaluation. As one of the most-widely used classifiers, the Multilayer Perception (MLP) has shown impressive classification performance. However, the MLP is very sensitive to the setting of the training parameters such as weights and biases. The traditional parameter training methods, such as, error back propagation algorithm (BP), are easily trapped into local optima and suffer premature convergence. To address these problems, this paper introduces a modified gravitational search algorithm (MGSA) by employing a multi-population strategy to let four sub-populations explore the different areas in search space and a Gaussian mutation operator to mutate the global best individual when swarm stagnate. After that, MGSA is used to optimize the weights and biases of MLP. The experimental results on a public dataset have validated the higher classification accuracy of the proposed method.

With the fast development of remote sensing techniques, the spatial resolution of remote sensed image are improved significantly. However, the excessive spatial resolution leads to a sharp increase in data volume and spectral information confusion of objects. The multi-scale kernel learning (MSKL) method has shown an excellent advantage in classification of high-resolution satellite image. Nevertheless, the performance of the MSKL is dramatically influenced by the widths and weights of the Radial Basis Function (RBF) kernel, since its multi-scale kernel function is constructed by several RBF kernels. In order to achieve efficient multi-scale classifier, a new dynamic differential evolution (DE) algorithm is introduced in this paper. In addition, the spectral features and spatial fractal texture features of images are synthetically employed to construct the multi-scale kernel. The experimental results show that the multi-scale kernel based on the dynamic DE algorithm is superior to the traditional multi-scale kernel in obtaining a better multi-scale kernel classifier and with higher classification accuracy.

Post-earthquake high resolution (HR) remote sensing image classification is crucial for disaster assessment and emergency rescue. 3-D convolutional neural networks (3-D CNNs) exhibit promising performance in remote sensing image classification. However, 3-D CNNs lack the theoretical underpinnings to perform multiresolution approximation for filter learning in view of the scale variance of natural objects. Gabor filtering can effectively extract multiresolution spatial information including edges and textures, which have a potential to reinforce the robustness of learned features in 3-D CNNs against the orientation and scale changes. In this paper, we propose a combined 3-D convolutional neural network and Gabor filters (GNN) method for post-earthquake HR image classification. Instead of choosing a single scale, GNN extends the spatial information to several scales by Gabor filters to take advantage of correlations among multiple scales for damage mapping. The experimental results show that GNN can reflect the multiscale information of complex scenes, obtain good classification results for mapping post-earthquake damage using HR remote sensing images.