基于多分支空谱特征增强的高光谱图像分类  

Hyperspectral image classification based on multi-branch spatial-spectral feature enhancement

作　　者：李铁[1] 李文许 王军国 高乔裕 LI Tie;LI Wenxu;WANG Junguo;GAO Qiaoyu(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105

出　　处：《液晶与显示》2024年第6期844-855,共12页Chinese Journal of Liquid Crystals and Displays

基　　金：辽宁省科技厅自然科学基金面上项目(No.2023-MS-314);辽宁省教育厅科学研究经费项目(No.LJ2020JCL007);辽宁省教育厅基本科研面上项目(No.LJKMZ20220678,No.LJKZ0357)。

摘　　要：为了解决高光谱图像自身及分类过程中噪声干扰大、空间-光谱特征信息提取不足以及有限样本下分类性能不佳等问题,提出一种基于多分支空谱特征增强的高光谱图像分类模型SSFE-MBACNN。首先,利用多分支特征提取模块分别提取浅层空谱特征和深层空间特征信息,并引入注意力机制抑制噪声干扰。其次,设计一种改进多尺度空谱特征提取融合模块及结合双池化和空洞卷积的空间特征增强模块实现空谱特征增强,减少模型参数量和提高分类性能。最后,用全局平均池化层代替全连接层,进一步降低参数量,缓解模型过拟合问题。实验结果表明,在Indian Pines(10%训练样本)、Pavia University (5%训练样本)和Salinas(1%训练样本)数据集分别取得了0.990 7、0.997 5和0.994 7的总体分类精度。SSFE-MBACNN不仅能充分利用空谱特征信息,而且在有限样本下也取得了优秀的分类性能,明显高于其他对比方法。To solve the problems of high noise interference in the hyperspectral image itself and the process of classification,insufficient extraction of spatial-spectral feature information,and poor classification performance under limited samples,a hyperspectral image classification model SSFE-MBACNN based on multi-branched spatial-spectral feature enhancement is proposed.First,shallow spatial-spectral feature information and deep spatial feature information are extracted separately using multi-branch feature extraction modules,and attention mechanism are introduced to suppress noise interference.Second,an improved fusion module for multi-scale spatial-spectral feature extraction and a spatial feature enhancement module combining dual pooling and dilated convolution are designed to achieve spatial-spectral feature enhancement,reduce the number of model parameters and improve classification performance.Finally,the global average pooling layer is used instead of the fully connected layer to further reduce the number of parameters and alleviate the model overfitting problem.The experimental results show that the overall classification accuracies of 0.9907,0.9975 and 0.9947 are achieved for the Indian Pines(10%training sample),Pavia University(5%training sample)and Salinas(1%training sample)datasets.SSFE-MBACNN makes full use of the spatialspectral feature information and achieves excellent classification performance with limited samples,which is significantly higher than other comparative methods.

关 键 词：高光谱图像分类 特征增强 多分支特征提取 注意力机制 多尺度特征 双池化 空洞卷积 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]