Toward Effective Hyperspectral Image Classification Using Dual-Level Deep Spatial Manifold Representation

摘要: Hyperspectral image (HSI) contains an abundant spatial structure that can be embedded into feature extraction (FE) or classifier (CL) components for pixelwise classification enhancement. Although some existing works have exploited simple structures (e.g., local similarity) to enhance either the FE CL component, few of them consider latent manifold and how simultaneously embed both seamlessly. Thus, their performance is still limited, especially in cases with limited noisy training samples. To solve problems one stone, we present a novel dual-level deep representation (SMR) network HSI classification, which consists two kinds blocks: SMR-based block block. In blocks, graph convolution utilized adaptively model lying each area. The difference former condenses SMR space form center pixel, while later leverages propagate label information other pixels within area one. train well, impose unsupervised loss on unlabeled samples supervised cross-entropy labeled joint learning, empowers utilize sufficient learning. Extensive experiments benchmark data set demonstrate efficacy proposed method terms