基于文本-光谱特征联合学习的高光谱图像分类算法

Hyperspectral Image Classification Algorithm Based on Textual-Spectral Feature Joint Learning

高光谱图像分类任务是遥感对地观测领域中的重要研究课题之一。针对高光谱图像覆盖范围广、地物种类多、人工标记难度高等问题,设计了一种基于文本-光谱联合学习的分类算法,利用文本模态的语义先验来增强不同场景之间的知识迁移能力,借助特征重建的方式学习判别和迁移信息,并采用自适应的文本嵌入交互模块挖掘编码器的潜在特征,实现了多模态特征之间的联合优化与分类效果提升。同时,采用4种不同算法进行对比验证,结果表明,新算法在单类别精度、总体精度(Overall Accuracy,OA)和Kappa系数方面均优于其他算法。

Hyperspectral image classification is one of the most important topics in earth observation.In order to tackle the challenges of hyperspectral images with wide coverage,diverse ground features,and high difficulty in manual labeling,a hyperspectral image classification algorithm based on joint learning of textual-spectral features was desighed.The algorithm leverages the semantic priors of the textual modality to enhance the knowledge transfer capability across different scenes,utilized feature reconstruction to learn discriminative and transferable information,and employed an adaptive textual-embedding interactive learning module to explore the latent features of the encoder,which finally achieved joint optimization of multi-modal features and improves classification performance.At the same time,four different algorithms were used for comparison and verification,and the results showed that the new algorithm was superior to other algorithms in terms of single-class accuracy,overall accuracy(OA)and Kappa coefficient.