基于层级Transformer的高光谱图像分类方法

Hyperspectral image classification method based on hierarchical transformer network

高光谱图像分类将每个像素分类至预设的地物类别,对环境测绘等各类地球科学任务有着至关重要的意义。近年来,学者们尝试利用深度学习框架进行高光谱图像分类,取得了令人满意的效果。然而这些方法在光谱特征的提取上仍存在一定缺陷。本文提出一个基于自注意力机制的层级融合高光谱图像分类框架(hierarchical self-attention network,HSAN)。首先,构建跳层自注意力模块进行特征学习,利用Transformer结构中的自注意力机制捕获上下文信息,增强有效信息贡献。然后,设计层级融合方式,进一步缓解特征学习过程中的有效信息损失,增强各层级特征联动。在Pavia University及Houston2013数据集上的试验表明,本文提出的框架相较于其他高光谱图像分类框架具有更好的分类性能。

Hyperspectral image classification,which assigns each pixel to predefined land cover categories,is of crucial importance in various Earth science tasks such as environmental mapping and other related fields.In recent years,scholars have attempted to utilize deep learning frameworks for hyperspectral image classification and achieved satisfactory results.However,these methods still have certain deficiencies in extracting spectral features.This paper proposes a hierarchical self-attention network(HSAN)for hyperspectral image classification based on the self-attention mechanism.Firstly,a skip-layer self-attention module is constructed for feature learning,leveraging the self-attention mechanism of Transformer to capture contextual information and enhance the contribution of relevant information.Secondly,a hierarchical fusion method is designed to further alleviate the loss of relevant information during the feature learning process and enhance the interplay of features at different hierarchical levels.Experimental results on the Pavia University and Houston2013 datasets demonstrate that the proposed framework outperforms other state-of-the-art hyperspectral image classification frameworks.