面向海岸带监测的高光谱图像深层空谱特征提取与分类

Hyperspectral remote sensing is an important tool for coastal zone observation. It has variant and significant applications in many fields such as environmental protection, resource utilization, and disaster detection. However, due to the high dynamicity, complexity, and diversity in coastal zone, the monitoring of these areas faces lots of novel challenges, such as the lack of training samples, and the low robustness of the classification. In order to solve these problems, this project investigates into deep-level spectral-spatial feature extraction and classification methods for land cover classification in coastal scene. On one hand, supervised spectral-spatial feature extraction methods will be developed based on the mechanism of visual perception, so as to capture the deep spatial correlation among hyperspectral pixels. On the other hand, active and semisupervised learning will be collaborated to enlarge the initial training set. Furthermore, a parallel and three dimensional deep learning framework will be developed for the classification of hyperspectral images. The execution of this project can promote the application of hyperspectral remote sensing in coastal zone monitoring and provide technical support for the decision making of the government, which satisfies the national construction strategy of China, and thus, has both important theoretical and applied values.

高光谱遥感是海岸带监测的重要手段，在近海环境保护、资源利用以及灾害检测等领域具有重要的应用价值。然而，由于海岸带地区具有高度的动态性、复杂性和多样性，海岸带高光谱对地监测面临训练样本缺失、分类方法鲁棒性低等一系列技术挑战。为解决复杂场景下海岸带地物分类的难题，本项目将深入研究高光谱图像深层空谱特征提取与分类方法。一方面，通过深入分析视觉感知机理，充分利用图像先验知识，挖掘高光谱图像中的深层空间结构关系，建立能有效描述高光谱图像的监督空谱结构特征模型。另一方面，通过设计基于协同学习的训练集优化策略、针对高光谱图像的三维深度学习网络和并行训练算法，实现高效精确的海岸带遥感监测。本项目的开展能够有效提升海岸带高光谱遥感数据的利用水平，为国家有关部门更高效精确的决策提供技术支撑，符合国家建设海洋强国的战略需求，具有重要的理论和应用价值。

湿地监测是生态环境保护的重要环节，传统湿地监测手段面临监测成本高、动态监测难、精测精度低等难题，高光谱遥感为湿地的精准监测提供了有效的途径。项目组提出了高光谱遥感图像空谱结构特征提取方法，构建了知识引导与数据驱动相结合的湿地遥感图像分类框架，为湿地地物的快速检测与精确识别提供了技术支撑。研究成果发表SCI论文17篇。研发的湿地遥感图像特征提取与分类技术，成功应用于黄河口湿地植被精细化制图等重要领域。