基于Transformer模型的高光谱图像分类算法研究

Research on Hyperspectral Image Classification Algorithm Based on Transformer Model

摘要: 高光谱图像(Hyperspectral image, HSI)分类在遥感领域扮演着关键角色。然而，处理高光谱图像分类任务时，遇到了光谱相同但物质不同、光谱不同但物质相同的复杂情况。尽管现有基于卷积神经网络(Convolutional Neural Network, CNN)的方法在局部信息处理方面表现出色，但它们在表示能力上存在一定限制。为了应对这一挑战，本文提出了一种综合考虑光谱信息和空间信息的Transformer方法(Spatial Spectral Transformer Network, SSTN)，即本文引入了Transformer结构，在光谱和空间中倾向于捕捉全局信息。通过构建Transformer，模型将空间光谱特征有机结合。通过规范的实验研究，本文发现：在IndianPines和Houston2013数据集的分类任务中，本文的方法相较于其它Transformer网络表现更为优越，并在与其他骨干网络的对比中具有显著的改进。

Abstract: Hyperspectral image (HSI) classification plays a crucial role in the field of remote sensing. When dealing with HSI classification tasks, there are complex situations where the spectra are the same but the substances are different, and the spectra are different but the substances are the same. Although existing methods based on Convolutional Neural Networks (CNN) perform well in local in-formation processing, they have certain limitations in representation ability. To address this chal-lenge, a Transformer method that comprehensively considers spectral and spatial information (Spatial Spectral Transformer Network, SSTN) has been proposed. To overcome this limitation, this article introduces a Transformer structure that tends to capture global information in both spec-trum and space. By constructing a Transformer, the model organically combines spatial spectral features. The rigorous experimental results show that in the classification task of the Indian Pines and Houston2013 datasets, our method performs better than other Transformer networks and shows significant improvement compared to other backbone networks.