基于改进反馈卷积自编码器的高光谱图像降维  

Hyperspectral image dimensionality reduction based on improved feedback convolutional auto-encoder

作　　者：刘芳华 宋文杰 LIU Fanghua;SONG Wenjie(Zhengzhou University of Light Industry,Zhengzhou 450002,China)

机构地区：[1]郑州轻工业大学,河南郑州450002

出　　处：《现代电子技术》2024年第19期94-99,共6页Modern Electronics Technique

基　　金：河南省科技攻关项目(242102210013);河南省高等学校青年骨干教师培养计划项目(2023GGJS090);国家自然科学基金项目(61502435)。

摘　　要：高光谱图像具有光谱分辨率高、特征丰富、图谱合一等优势,在土地利用分类、城市规划管理、森林资源调查等领域有着广泛应用。但是,高光谱图像不同光谱通道间存在大量冗余信息,导致高光谱图像降维算法复杂度高,同时也降低了高光谱图像降维算法的性能。针对该问题,结合目前主流的深度学习技术,文中提出一种基于改进反馈卷积自编码器的高光谱图像降维方法。首先,为增强信息的流动,在原有编码器模型中引入残差连接,促进了梯度信息的传播;其次,为了能够更好地捕捉高光谱数据的关键特征,在现有解码器模型中加入分支结构;最后,通过使用平均池化替换最大池化,采用平均绝对误差(MAE)替换均方误差(MSE)损失函数,进一步优化模型的特征提取能力,提高高光谱图像降维的性能。实验结果表明,所提出模型在Indian Pines数据集上的高光谱图像降维性能超过了现有最新方法,为高光谱图像降维提供了一个新思路。Hyperspectral images(HSIs)are characterized by high spectral resolution,abundant spectral features,imagery and spectral integration,and have been widely utilized in various fields,such as land utilization classification,urban planning and management,and forest resource investigation.However,the presence of a large amount of redundant information among different spectral channels in HSIs results in the high complexity of HSI dimensionality reduction algorithms and decreases its performance.In view of the above,an HSI dimensionality reduction method based on improved feedback convolutional autoencoder is proposed in combination with the existing mainstream deep learning technology.To facilitate the information flows,a residual connection is introduced into the original encoder model,which promotes the gradient information propagation.To better capture the key features of the HSI data,a branching structure is incorporated into the original decoder model.The maximum pooling is replaced with the average pooling,and the mean absolute error(MAE)is used to replace the mean square error(MSE)loss function,so as to further optimize the feature extraction ability of the model and improve the performance of HSI dimensionality reduction.The experimental results show that the proposed model outperforms the latest methods in terms of HSI dimensionality reduction on the Indian Pines dataset,so it provides a new idea for HSI dimensionality reduction.

关 键 词：高光谱图像 高光谱图像降维 反馈卷积 自编码器 深度学习 分支结构 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP39[电子电信—信息与通信工程]