The fusion of spectral-spatial features based on deep learning has become the focus of research in hyperspectral image (HSI) classification. However, previous deep frameworks based on spectral-spatial fusion usually performed feature aggregation only at the branch ends. Furthermore, only first-order statistical features are considered in the fusion process, which is not conducive to improving the discrimination of spectral-spatial features. This article proposes a global-local hierarchical weighted fusion endto-end classification architecture. The architecture includes two subnetworks for spectral classification and spatial classification. For the spectral subnetwork, two band-grouping strategies are designed, and bidirectional long short-term memory is used to capture spectral context information from global to local perspectives. For the spatial subnetwork, a pooling strategy based on local attention is combined to construct a global-local pooling fusion module to enhance the discriminability of spatial features learned by a convolutional neural network. For the fusion stage, a hierarchical weighting fusion mechanism is developed to obtain the nonlinear relationship between both spectral and spatial features. The experimental results on four real HSI datasets and a GF-5 satellite dataset demonstrate that the method proposed is more competitive in terms of accuracy and generalization.

No abstract

Convolutional neural network (CNN) is widely used in HSI classification owing to their advantages of spatial-spectral features capture capability, learning depth features as well as their structural flexibility. Nevertheless, the shape of the convolution kernel is fixed, a limitation that leads to shape fixation when modeling different feature in CNN, especially in the edge regions between classes. A multiscale adaptive convolution (MSAC) model is proposed in this paper to overcome this shortcoming. Combining superpixels with the traditional convolutional kernels to form adaptive kernels automatically adjusts the receptive field, suppresses edge noise, and enhances feature learning for different classes. On the basis of adaptive convolution, adaptive convolution units (AConvUs) are constructed. The hierarchical residual structure is constructed by superimposing multiple AConvUs to learn the spatial spectral features of different receptive fields of the HSI, reduce the gradient disappearance and enhance the robustness. The proposed multiscale convolution adjusts the shape of the convolution kernel according to the spatial distribution of different superpixels in HSI. Finally, the MSAC classification framework is constructed by the decision fusion of multiscale adaptive convolution at different superpixel scales, which helps to extract the complementary information of the HSI. The MSAC method's experimental performance on several HSI datasets, including the Gaofen-5 (GF-5), to verify the validity and practicality.

No abstract

No abstract