首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 918 全文下载次数: 1242
引用本文:

DOI:

10.11834/jrs.20210550

收稿日期:

2020-12-07

修改日期:

2021-07-01

PDF Free   EndNote   BibTeX
生成对抗网络的无监督高光谱解混
靳淇文1, 马泳1, 樊凡1, 黄珺1, 李皞2, 梅晓光1
1.武汉大学电子信息学院;2.武汉轻工大学数学与计算机学院
摘要:

近年来,基于深度学习中自编码器(AE)的方法在无监督的高光谱解混中受到了广泛的关注。由于AE的学习过程可以表述为通过训练找出一组低维的隐藏层(丰度),并用其对应的权重(端元)进行组合来减少重构误差,因而这种框架被广泛迁移并应用于高光谱的解混算法之中。然而,现有基于AE的框架虽然能有效地处理无监督的解混场景,却都存在着对噪声和初始化条件不鲁棒的问题,且解混精度也有待进一步提升。针对以上问题本文提出了一种全新的基于对抗性自编码网络(AAE)的无监督解混网络框架。首先,我们在网络的生成器中根据丰度和为一(ASC)及非负性(ANC)的物理意义,设计了一个基于AE的端到端解混框架。然后,在网络的判决器中本文采用初始化的丰度图作为真实值,将生成器的隐藏层(丰度)与初始化的丰度进行对抗训练,在重构误差与对抗误差的同步优化中提升框架解混性能。与传统的AE方法相比,该方法通过引入对抗性过程,在判决器中加入丰度的先验知识,可以大大提高框架的性能和鲁棒性。仿真和真实的高光谱数据的实验表明,该算法较现有方法相比具有更高的解混精度。

Hyperspectral Unmixing Based on Adversarial Autoencoder Network
Abstract:

Due to the limitation of spatial resolution of instruments and complex natural surfaces, spectral mixing (SU), which refers to identify the proportion (abundance) of the basic component spectrum (endmember) in the sub-pixel level, has become a significant issue for the deep development of hyperspectral image analysis. Since the training procedure can be explained as finding a set of low-dimensional representations (abundance) that reconstruct the data with their corresponding bases (endmembers), the autoencoder (AE) based method has received much attention in the unsupervised hyperspectral unmixing. However, although the existing AE methods can effectively deal with unsupervised scenarios of unmixing, the performance has not been satisfactory, noise and initialization conditions will greatly affect the performance of unmixing. In this paper, we proposed a novel technique network for unsupervised unmixing which is based on the adversarial autoencoder network, termed as AAENet. First, the generator is designed as an end-to-end unmixing network based on AE to obtain the meaningful abundance subjected to abundance nonnegative constraint (ANC) and abundance sum-to-one constraint (ASC). Then, we take the adversarial training process to map the abundance prior to the hidden code vector (abundance), which is equivalent to providing an adaptive training error to correct the AAENet converging towards a more accurate and interpretable unmixing solution. The experiments on simulated and real hyperspectral data (jasper dataset) both demonstrate that the proposed algorithm can outperform the other state-of-the-art methods. The synthetic data is polluted by Gaussian noise at different levels, where the SNR is varied from 10 to 30dB with an interval of 10dB. Each algorithm is run 10 times and the average and standard deviation are reported. With the increase of the noise level, the proposed algorithm exhibits much better robustness in both the abundance and endmember estimation and achieves the best or comparable results in all cases. For real dataset experiments, the results of AAENet not only showed sparse abundances for the region but also interpreted the boundary as a combination of neighboring materials. The best results are still obtained in overall highly mixed scenes. Compared with the traditional AE method, our approach can greatly enhance the performance and robustness of the model by using the adversarial procedure and incorporating the abundance prior to the framework. The discrimination network is designed to make it possible to transfer the potentially intrinsic properties of the abundance prior information into consideration. The main purpose of our method is to take an adversarial training process to impose a prior on the hidden code vector of the autoencoder, and guide the output of the hidden code vector results in meaningful samples. The experiments on simulated and real hyperspectral data demonstrate that the proposed algorithm can achieve better unmixing performance compared with state-of-the-art methods.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群