摘要:
高光谱图像数据具有维度高、数据稀疏、空间光谱信息丰富等特点,针对空谱联合分类模型中高光谱图像卷积操作处理大片相同类别像素区域时会存在计算的空间冗余,3D卷积对深层空间纹理特征提取不充分,串行注意力机制结构不能充分考虑空谱相关性的问题,该文提出了改进的3D-Octave卷积高光谱图像分类模型。首先改进的3D-Octave卷积模块将输入的高光谱图像数据划分为高频特征图和低频特征图,减少空间信息冗余,提取多尺度的空间光谱特征,结合跨层融合策略,加强对浅层空间纹理特征和光谱特征的提取;随后利用2D卷积提取深层空间纹理特征并进行光谱特征融合;最后使用三维注意力机制跨纬度交互实现对有效特征的关注和激活,增强网络模型的性能和鲁棒性。结果表明,由于充分提取有效空谱联合特征,在印第安松树林(Indian Pines, IP)数据集的训练集比例为10%的条件下,OA,Kappa和AA分别为99.32%,99.13%和99.15%; 在帕维亚大学(Pavia University, PU)数据集的训练集比例为3%的条件下,OA,Kappa和AA分别为99.61%, 99.44%和99.08%。与5个主流分类模型进行对比,获得了更高的分类精度。
Abstract:
Hyperspectral image data are characterized by high dimensionality, sparse data, and rich spatial and spectral information. In spatial-spectral joint classification models, convolution operations for hyperspectral images can lead to computational spatial redundancy when processing large regions of pixels of the same category. Furthermore, the 3D convolution fails to sufficiently extract the deep spatial texture features, and the serial attention mechanism cannot fully account for spatial-spectral correlations. This study proposed an improved 3D Octave convolution-based model for hyperspectral image classification. First, the input hyperspectral images were divided into high- and low-frequency feature maps using an improved 3D Octave convolution module to reduce spatial redundancy information and extract multi-scale spatial-spectral features. Concurrently, a cross-layer fusion strategy was introduced to enhance the extraction of shallow spatial texture features and spectral features. Subsequently, 2D convolution was used to extract deep spatial texture features and perform spectral feature fusion. Finally, a 3D attention mechanism was used to focus on and activate effective features through interactions across latitudes, thereby enhancing the performance and robustness of the network model. The results indicate that, due to the adequate extraction of effective spatial-spectral joint features, the overall accuracy (OA), Kappa coefficient, and average accuracy (AA) were 99.32%, 99.13%, and 99.15%, respectively in the case where the Indian Pines (IP) dataset accounted for 10% in the training set and were 99.61%, 99.44%, and 99.08%, respectively when the Pavia University (PU) dataset represented for 3% of the training set. Compared to five mainstream classification models, the proposed method exhibits higher classification accuracy.
下载: