面向深空应用的高光谱图像高精度分类方法研究

Hyperspectral imaging technology can detect hundreds of spectral channels, which can collect data from both of the spatial and spectral domains. With the advantage of high spectral resolution, hyperspectral imager has already become a necessary load in deep space exploration. However, in practical applications, due to the limitations of several conditions such as the detection storage and transmission in deep space exploration, the hyperspectral data has several typical features, such as low spatial resolution, scarcity of the target sample and unknown category, etc. All these characteristics lead a tough problem that hyperspectral classification is too difficult to achieve high classification accuracy, which is far from the desired scientific and engineering objectives. Therefore, to address the aforementioned issues, this project mainly promotes a research route of "deep feature enhancement-nonlinear sample generation-high precision classification" and develops an efficient and robust high precision classification method. First of all, a feature enhancement model based on deep spatial-spectral double residuals network is built to enhance the representation characteristics of data. Secondly, a nonlinear sample generation model based on generative adversarial networks and the orthogonal complement space projection is designed to provide reliable data conditions. Finally, an efficient classifier based on deep kernel extreme learning and prototype learning theory is established to achieve high accuracy and low complexity classification. This project can provide effective technical support for observing and classifying the surface environment of the stars. Especially, under the premise of the national Mars exploration project, it can provide an important prospective basic research for the development and utilization of space resources for the country, and can provide a solid theoretical foundation for the on-board realization of intelligent algorithms in the follow-up deep space exploration, which has extremely significant application value.

高光谱成像技术可探测成百上千个光谱通道，具有图谱合一和光谱分辨率高的优势，已成为深空探测的必备载荷。但在实际应用中，受探测资源及传输等条件的限制，深空高光谱图像存在空间分辨率低、目标样本稀缺、未知类别等问题，导致分类难度大精度低，与期望的科学和工程目标存在较大差距。本项目拟提出“深度特征增强-非线性样本生成-高精度分类”的研究思路，发展高效、鲁棒的高精度分类方法。（1）建立基于深度空谱双残差网络的特征增强模型，增强数据的表示特性；（2）设计基于生成对抗网络和正交补空间投影的非线性样本生成模型，提供可靠的数据条件；（3）建立协同深度核极限学习与原型学习理论的高效分类器，实现高精度低复杂度的分类。可为星体表面环境观测和物质分类提供有效的理论和技术支撑；在火星探测计划立项的前提下，为国家进行资源开发及利用提供前瞻性的基础研究；为深空探测智能算法在轨实现提供相关的基础技术，具有极其重要的应用价值。

在实际应用中，受探测资源及传输等条件的限制，深空高光谱图像存在空间分辨率低、目标样本稀缺、未知类别等问题，导致分类难度大精度低，与期望的科学和工程目标存在较大差距。本项目的主要研究内容包括构建了基于深度空谱双残差网络的特征增强模型，增强数据的表示特性；设计了基于生成对抗网络和正交补空间投影的非线性样本生成模型，提供可靠的数据条件；构建了协同深度核极限学习与原型学习理论的高效分类器，实现了高精度低复杂度的高光谱图像分类。该项目共发表学术论文X篇，其中SCI检索论文X篇，申请发明专利X项，软件著作权1项。所提出的方法，可为星体表面环境观测和物质分类提供有效的理论和技术支撑；为国家进行资源开发及利用提供前瞻性的基础研究；为深空探测智能算法在轨实现提供相关的基础技术，具有极其重要的应用价值。特别地，这些方法还可应用在传统高光谱图像具有稀少目标样本的特定应用中，例如：军事伪目标检测、农作物病虫害范围估计等。