基于改进DenseNet和空谱注意力机制的高光谱图像分类  

Hyperspectral Image Classification Based on Modified DenseNet and Spatial Spectrum Attention Mechanism

作　　者：王欣[1] 樊彦国[1] Wang Xin;Fan Yanguo(College of Oceanography and Spatial Information,China University of Petroleum(East China),Qingdao,Shandong 266500,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266500

出　　处：《激光与光电子学进展》2022年第2期203-214,共12页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(62071492);山东省重点研发计划(2019GHY112017)。

摘　　要：针对高光谱图像维度高、训练样本少以及训练带来的过拟合、参数过多问题,提出了一种基于改进密集连接网络(DenseNet)和空谱注意力机制网络(MDSSAN)的高光谱图像分类方法。首先,对高光谱图像进行主成分分析,并将中心像素的空间邻域输入改进的网络模型中。然后,对三维DenseNet进行改进,将模型中的三维卷积块分解成空间维和光谱维的采样卷积。最后,在空间维上引入空间注意力机制,在光谱维上引入通道注意力机制,以减少模型的训练参数,提取更具判别力的空谱联合特征。实验结果表明,MDSSAN模型在Indian Pines、Pavia University和KSC数据集上的总体分类精度分别为99.43%、99.74%、98.98%,相比其他对比模型,该模型的收敛速度更快,分类性能更好。Aiming at the problems of hyperspectral image with high dimension, a few training samples, over fitting and too many training parameters, we propose an modified dense connection network(DenseNet) combined with spatial spectrum attention mechanism network(MDSSAN). First, the hyperspectral images are analyzed by principal component analysis, and the spatial neighborhoods of the central pixels are input into the modified network model. Then, three-dimensional DenseNet is improved, and the three-dimensional convolution block in the model is decomposed into the sampling convolution of the spatial dimension and the spectral dimension. Finally, the spatial attention mechanism is introduced in the spatial dimension, and the channel attention mechanism is introduced in the spectral dimension to reduce the training parameters of the model and extract more discriminative space-spectrum joint features. Experimental results show that the overall classification accuracy of the MDSSAN model on the Indian Pines, Pavia University, and KSC data sets are 99.43%, 99.74%, and 98.98%, respectively. Compared with other comparison models, the model has faster convergence speed and better classification performance.

关 键 词：图像处理 高光谱图像 密集连接网络 空间注意力机制 通道注意力机制 

分 类 号：P407.8[天文地球—大气科学及气象学]