基于对抗网络的高光谱图像半监督分类方法研究

Due to the effects of multipath scattering and mixed pixels, hyperspectral data often have nonlinear distribution properties. Deep learning can automatically learn the intrinsic features from nonlinearly distributed data in a hierarchical manner, thus achieving more and more attention in the field of hyperspectral image classification. However, most of the existing methods are based on supervised learning models. Their performances depend on the quantity and quality of training samples, while it is very difficult to acquire the label information of hyperspectral images. Besides, hyperspectral data are easily corrupted by multiple kinds of noises in the collection process, which badly affect the classification performance. To address these key issues, this project plans to study a robust semi-supervised classification model, aiming at sufficiently taking advantage of unlabeled samples and improving the classification performance of hyperspectral images. The main research contents of this project include: a hyperspectral image denoising model based on cascaded convolutional neural networks; a spectral-spatial feature representation model using joint spectral and spatial attention modules; and a semi-supervised model for hyperspectral image classification using a multi-task generative adversarial network. The project is expected to provide new research ideas and methods for hyperspectral data analysis.

由于多径散射效应、混合像元等的影响，高光谱数据通常具有非线性分布特性。深度学习算法通过多层抽象机制，可自主学习非线性分布数据的高层本质特征，因而在高光谱图像分类领域受到了越来越多的关注。但是，当前的方法大多基于监督学习模型，其分类性能依赖训练样本的数量和质量，而高光谱图像的标注往往很难获取。此外，高光谱数据在采集过程中容易受多种不同噪声干扰，这些噪声严重影响了分类精度和可信度。本项目针对这些问题，拟研究鲁棒的半监督分类方法，充分利用未标注样本的信息，提升高光谱图像分类性能。主要研究内容包括基于级联卷积网络的高光谱混合噪声去除模型、联合光谱和空间注意力机制的谱-空特征表示模型、以及基于多任务生成对抗网络的高光谱图像半监督分类模型。本项目的开展将为高光谱数据分析提供新的研究思路和方法。

深度学习可自主学习非线性分布数据的高层语义特征，因而在高光谱图像分类领域受到了越来越多的关注。但是，当前的方法大多基于监督学习模型，其分类性能依赖训练样本的数量和质量。为此，本项目首先研究了高光谱图像重建算法，在先验信息引导下，可直接从RGB图像重建高光谱图像，从而有效地克服了高光谱图像获取困难的问题；在此基础上，研究了高光谱图像的光谱和空间特征表示模型，通过注意力机制有效地提升了特征的判别能力；其次，为了克服标注数据少的问题，着重研究了基于多任务生成对抗网络的半监督特征提取以及基于跨模态对比学习的无监督特征提取模型；最后，探索了高光谱图像和激光雷达数据融合模型，进一步提升了分类性能。相关研究成果共发表论文16篇，其中CCF A类和IEEE Transactions论文12篇，共有4篇论文先后入选ESI高被引论文。此外，授权发明专利2项，培养博硕士研究生6名，完成了预期的研究目标。