基于自适应上下文聚合网络的双高遥感影像分类

Adaptive context aggregation network for H^(2) remote sensing imagery classification

融合高光谱和高空间分辨率(双高)遥感的优势可以实现地物目标更为全面和精细的属性识别。然而,空间分辨率的显著提升使得双高影像中地物细节特征凸显出来,呈现出极高的空谱异质性,进而导致同物异谱现象大量发生,地物类内方差明显增大。基于此,本文提出一种局部-全局上下文信息自适应聚合的快速双高影像分类框架(adaptive context aggregation network,ACANet),通过编码-解码的全卷积网络架构顾及全局空谱信息,在编码器中构建局部到全局的长距离上下文感知模块缓解双高影像极大的类内方差,在解码器中构建自适应上下文聚合模块进一步实现局部和全局的上下文信息自适应聚合。本文方法在WHU-Hi双高影像分类基准数据集中取得了优异的分类性能,试验表明可以很好缓解双高影像极高空谱异质性对地物精细分类的影响。

High spectral and spatial resolution(H^(2))remote sensing imagery can achieve more comprehensive and precise attribute recognition of ground objects.However,the details of ground objects are gradually revealed with the significant improvement of the spatial resolution,which makes the H^(2) images show extremely high spectral variability and spatial heterogeneity,and then the phenomenon of the same class with different spectrums occurs in large numbers;the intraclass variance increases significantly.As a result,an adaptive aggregation context network was proposed for H^(2) image classification,which uses a full convolution network with encoder-decoder architecture to achieve global spectrum-spatial fusion.A local-to-global long-distance context module was designed to alleviate the intraclass variance in the encoder module.Then an adaptive context aggregation module was constructed in the decoder module for the adaptive aggregation of local and global context information.ACANet has achieved excellent performance in the WHU-Hi benchmark dataset,and the experiments show that it can sufficiently alleviate the impact of the spatial-spectrum heterogeneity of the H^(2) image in the precise classification.