基于深度流形蒸馏网络的高光谱遥感图像场景分类方法

Hyperspectral remote sensing image scene classification method based on deep manifold distillation network

当前场景分类任务大多面向高分辨率遥感图像,由于缺乏光谱信息限制了它的场景鉴别能力,而高光谱遥感图像具有“空谱合一”的特性,在场景分类问题上具有独特优势。针对高光谱遥感图像中地物分布复杂,以及高光谱图像中维度高、存在冗余等问题,本文提出一种高光谱场景分类流形蒸馏网络(hyperspectral scene classification manifold distillation network,HSCMDNet),有效提高了分类性能。对于遥感图像地物分布复杂问题,HSCMDNet模型使用基于移位窗口的层次化视觉Transformer(hierarchical vision transformer using shifted windows,SwinT)作为教师网络来充分挖掘高光谱图像的长距离依赖信息,捕获不同波段之间的关系。在此基础上,在教师网络与ResNet-18学生网络之间设计流形蒸馏损失,通过在流形空间中匹配学生和教师的中间层输出特征实现教师模型的知识更有效地向轻量化学生模型转移,缓解了高光谱图像中维数高导致的高计算复杂性问题。在欧比特高光谱图像场景分类数据集(Orbita hyperspectral image scene classification dataset,OHID-SC)及天宫二号遥感图像自然场景分类数据集(natural scene classification with Tiangong-2 remotely sensed imagery,NaSC-TG2)上,所提出的HSCMDNet网络的最佳分类精度分别达到了93.60%和94.55%。

Most of the current scene classification tasks are based on high-resolution remote sensing images,and the lack of spectral information limits its discrimination ability for scene classification.While hyperspectral remote sensing images have the characteristic of“spatial-spectral integration”,which has unique advantages in scene classification.To address the issue of the complex landcover distribution and the high dimension and redundancy in hyperspectral images,this paper proposes a hyperspectral scene classification manifold distillation network(HSCMDNet)to improve the performance of hyperspectral remote sensing scene classification.For the complex landcover distribution of remote sensing images,HSCMDNet employs Swin Transformer as a teacher network to reveal the long-distance dependency information of hyperspectral images and capture the relationship between different bands.After that,a manifold distillation loss is designed in the middle layer of the teacher network and the student network ResNet-18.By matching the middle layer output features of the student model and the teacher model in the manifold space,the knowledge of the teacher model is transferred to the lightweight student model effectively,which alleviates the high computational complexity caused by high-dimensional hyperspectral data.In the orbita hyperspectral image scene classification dataset(OHID-SC)and natural scene classification with Tiangong-2 remotely sensed imagery(NaSC-TG2),the best classification accuracies of the proposed HSCMDNet network reached 93.60%and 94.55%,respectively.