基于轻量级光谱-空间注意力交互网络的高光谱地物分类研究

通过引入基于卷积神经网络(convolutional neural network, CNN)的分类算法，高光谱图像(hyperspectral image, HSI)分类任务的精度取得显著的提升，但目前主流CNN算法往往较为复杂且参数量大，从而导致网络难以训练以及容易产生过拟合问题。为在保证网络分类性能的前提下实现轻量化，本文提出一个轻量级架构的基于光谱-空间注意力交互机制的CNN网络用于HSI分类。为实现HSI的光谱-空间特征提取，构建了一个轻量化的双路径骨干网络用于两种特征的提取和融合。其次，为提高特征的表征能力，设计了两个注意力模块分别用于光谱和空间特征的权重再调整。同时，为加强双路径特征之间的关联以实现特征的更好融合，注意力交互机制被引入到网络中以进一步提升网络性能。在3个真实HSI数据集上的分类结果表明，本文所提网络可达到99.5%的分类准确度，并相比于其他网络至少减少50%的参数量。

Lightweight spectral-spatial attention interactive network for hyperspectral image classification

By introducing the convolutional neural network (CNN) based classification algorithms,the accuracy of hyperspectral image (HSI) classification task has been significantly improved. However,the present mainstream CNN algorithms are usually complex and have a large number of parameters,which will make network hard to train and cause the over-fitting problem.To achieve lightweight on the premise of ensuring the performance of network,we proposed a lightweight CNN network based on spectral-spatial attention interaction mechanism for HSI classification task.To realize the spectral and spatial feature extraction,we constructed a lightweight dual-path backbone network for the extraction and fusion of the two kinds of features.And,to improve the representation ability of features,we designed two attention modules to adjust the weight of spectral and spatial features respectively.At the same time,to strengthen the correlation between two path features to achieve better fusion,attention interaction mechanism was introduced into the network to further improve the network performance. The classification results on three real HSI data sets show that the proposed network can achieve 99.5% classification accuracy and reduce at least 50% parameter amount compared with other networks .