基于典型目标候选区域生成的高光谱遥感影像超分辨率算法研究

Fine spatial description is an important prerequisite for accurate interpretation of hyperspectral remote sensing images, and is also the key to promote its wide application. Both the limitation of hardware and the complex imaging environment result in low spatial resolution of hyperspectral remote sensing images, which is difficult to meet with the application requirements. Meanwhile, with the development of earth observation technology, the image size produced by remote sensing imaging system has increased significantly. With respect to the existing researches focused on super - resolution all contents in the image, which is time-consuming and the effectiveness for the target cannot be guaranteed, this project carries out the research on typical targets’ regional proposal based hyperspectral remote sensing images super-resolution methods: (1) With the combination of the data enhancement and so on technologies, and embedding the deconvolution layer in the feature extraction network, typical target feature modeling under multi-scale and multi-view is to be realized; (2) By analyzing the difference among bands, a feature enhancement model is to be built, and a multi-scale feature fusion based typical targets’ regional proposal neural network is to be constructed; (3) This project designs an spectral difference generative adversarial network, in which the noise in the generator is designed under the guidance of the feature model. This operation fast super-resolves the images with spectral preservation. The research results of this project will provide effective theoretical support for the extensive application and rapid development of hyperspectral remote sensing images.

精细的空间描述是准确解译高光谱遥感影像的重要前提，也是推进其应用广泛化的关键所在。硬件设备的限制以及复杂的成像环境，均造成高光谱遥感影像空间分辨率低，难以满足应用需求。同时，遥感技术的发展，使得成像系统产生的影像幅面显著增大。针对现有研究对影像中全部区域进行均一的超分辨率处理，造成目标区域的有效性无法保证且计算复杂度高的问题，本课题拟开展基于典型目标候选区域生成的高光谱遥感影像超分辨率算法研究：（1）结合数据增强等技术，并在特征提取网络中内嵌上采样层，实现多尺度、多视角下典型目标特征建模；（2）分析波段间的差异性，建立特征增强模型，构建基于多尺度特征融合的典型目标候选区域生成模型；（3）设计光谱差生成对抗网络模型，将目标特征模型指导生成器中的噪声设置，实现光谱保真前提下影像目标候选区域的快速有效超分辨率处理。本课题成果将为高光谱遥感影像在对地观测和深空探测领域的广泛应用提供有效的理论支持。

高光谱遥感影像由于其丰富的光谱信息，能够同时获取地物的空间以及光谱信息，这为准确、定量地分析地物特性提供了支撑，逐渐成为对地观测中的一种重要信息源。然而，由于成像设备和成像环境的限制，高光谱遥感图像空间分辨率低，成像幅宽大，从而导致用户同时获取观测目标高空间分辨率和光谱分辨率的需求无法满足。本项目围绕基于典型候选目标区域生成的高光谱遥感影像超分辨率算法开展研究，解决高光谱图像空间分辨率低的问题，主要研究内容包括：（1）典型遥感目标特征建模；（2）大幅面高光谱遥感影像典型目标候选区域生成算法；（3）高光谱遥感影像典型目标候选区域快速超分辨率算法。本项目的研究成果可以根据遥感影像中不同区域的具体内容，动态地进行超分辨率处理，有效提高遥感影像中目标候选区域地细节信息，强化特征，进一步促进了高光谱图像在城市规划、精细农业、灾害预警等领域地应用。目前，在该项目的资助下，项目组已发表科研论文19篇，其中SCI检索论文10篇，CCF B类中文期刊论文3篇，申请专利4项，其中授权1项，培养硕士研究生6名。