深层网络下高光谱遥感影像特征表达与融合

The high quality image feature plays a key role in the intelligent interpretation and automatic recognition of the hyperspectral remote sensing image. However, the traditional feature extraction methods often ignore the deep semantic information in the image, it will result in the lack of the expression ability, the learning ability and the adaptability for the derived features. With respect to the problem mentioned above, this project will introduce the deeper networks to conduct the feature extraction and fusion of hyperspectral remote sensing images. The main research includes: 1) develop a super-resolution technique for hyperspectral image under the deeper networks through optimizing the network and adding the inter-class and intra-class regularizations, it can highlight the clear edge features; 2) construct the framework of feature extraction under the deeper network with the help of transfer learning, it can break the bottleneck of training the deeper network with the limited training samples; 3) establish an end-to-end fusion strategy for spectral and spatial features using the two-stream architecture to improve the expression ability of the features. The purpose of this project is to extract the features with greater expression ability, learning ability and the adaptability, and to realize the theory and methods innovation for feature extraction and fusion using deeper network. It also aims at providing the technical support for the extensive application of hyperspectral remote sensing image.

高质量的影像特征对高光谱遥感影像智能解译、地物自动分类与识别起着关键作用。然而目前传统的特征提取方法往往忽略了影像深层语义信息，导致特征表达能力有限，学习能力和自适应能力较弱。针对上述问题，本项目拟引入深层深度学习网络进行高光谱遥感影像特征提取和融合，研究内容包括：1) 优化网络结构，构建深层网络下影像超分辨率模型，融入类间类内正则因子，提取边缘清晰的影像特征；2) 借助迁移学习技术，提出迁移深层网络下特征提取框架，突破深层网络在训练样本不足情况下难以训练的技术瓶颈；3) 利用双分支网络结构，建立端对端的空-谱特征的融合策略，提高特征表达能力。本项目旨在提取具有深层语义信息、学习能力及自适应能力强的特征，实现深层网络下特征提取和融合理论与方法的创新，为高光谱遥感影像的广泛应用提供技术支撑。

随着我国高分工程的加速建设，遥感影像的数据量正以每天TB级的速度增长，为后续应用提供了重要的数据保障，也对影像特征表达技术提出了更高的要求。然而，现阶段的遥感影像特征表达技术还不够成熟，这些都迫使我们建立一种自动、快速、高精度的高光谱遥感影像特征表达与融合框架对地表进行全面观测和综合分析，其有着海量的业务需求和广泛的应用前景。在项目执行期，构建了双分支深层结构下的高光谱图像分类方法、基于类损失函数的超分辨率高光谱影像分类方法、基于递归神经网络由粗到细的迭代式分类方法，相关成果均发表在IEEE Transactions on Geoscience and Remote Sensing期刊上，极大地推动了深度学习网络在高光谱遥感影像中的应用，有效地缓解了高光谱影像有标记样本过少造成的网络训练难的问题。同时，提出方法也可以应用于不同领域，例如将提出的基于递归神经网络由粗到细的迭代式分类方法应用于虚拟现实眼镜中，可以准确地进行人手的分割与定位，有助于人与虚拟物体进行交互，在VR眼镜中显著提升了用户体验；将提出的双分支的特征融合方法应用于岩性预测中，可以准确分析出岩石的组成。本项目有效地将理论研究延伸至实际应用中，产生了重要的市场价值。