面向高光谱图像分类的轻量级3D卷积网络模型构建与优化

Hyperspectral image(HSI) containing both spatial and spectral information, which brings a significant opportunity for land-cover classification and also poses a challenge for classification algorithm. Recent years, deep learning based HIS classification approaches have achieved significant breakthroughs. In deep learning based HSI classification method, a large number of labelled samples are needed for training and just very limited labelled samples are available in HSI, which has been a bottleneck to further improve the classification performance. .This project aims to solve these issues by building and optimizing deep network models to finish HSI classification task under the small sample condition. There will be four research components. First, a novel lightweight 3D deep convolutional neural network model which can fit the 3D structure of HSI will be developed. On the premise of accuracy, the scale of network (the number of learnable parameters and computation costing) will be decreased as much as possible to avoid over fitting. Second, in order to improve the network structure，an automatic model optimization approach that leverages reinforcement learning and knowledge transfer will be proposed. Third, a 3D twinborn network that combined with deep metric learning will be developed. For increasing distinguishability，the HSI will be transfer from image space to metric space via 3D twinborn and will be classified in metric space. Finally, a dual learning model that combine shallow clustering model and deep classification model will be studied. In dual learning model, the shallow model and deep model promote each other’s performance via dual learning and finish HSI classification in the situation of small sample..The research result will be expended in the field of theory and application of deep learning, which improves the level of data analysis and interpretation of remote sensing observation.

高光谱图像（HSI）同时包含了空间信息和光谱信息，给地物分类带来了巨大机遇，同时也给分类算法提出了挑战。近年来基于深度学习的HSI分类方法取得了突破性的进展。然而，深度学习模型对大量标注样本的需求和HSI小样本问题之间的矛盾，成了此方法取得进一步发展的瓶颈。本项目针对这些问题，研究面向HSI小样本分类的深度网络模型的构建及优化：构建适合HSI三维结构的轻量级3D深度卷积网络，在保证分类精度的情况下，尽可能减少网络规模（参数量和计算量），有效避免小样本下网络过拟合；结合增强学习和知识迁移，实现轻量级网络的自主学习和优化；构建结合深度度量学习的3D孪生网络模型，将HSI从图像空间映射到易于区分的相似性度量空间完成分类任务；构建结合浅层聚类模型和深层分类模型的对偶模型，通过对偶学习两个模型互相促进，实现小样本条件下的HSI空谱联合分类。成果将扩展深度学习的理论和应用，提高遥感观测数据解译水平。

高光谱图像分类是高光谱遥感对地观测技术的一项重要内容，在军事及民用领域都有着重要的应用。然而，高光谱图像的高维特性、波段间高度相关性、光谱混合、小样本等问题使得高光谱图像分类面临巨大挑战。本项目从高光谱图像分类面临的挑战问题出发，将机器学习和模式识别领域的最新技术与高光谱图像处理技术相结合，针对高光谱图像和地物目标的特点，研究面向小样本条件下的高光谱图像空谱联合分类的深度学习模型的构建与优化，研究内容包括适于高光谱图像特点的轻量级3D卷积网络的构建；基于可微分网络结构搜索的轻量化网络结构的自主优化；小样本条件下，基于迁移学习的轻量级3D卷积网络模型、结合深度度量学习的轻量级3D孪生网络分类模型以及联合轻量化网络和深度聚类的半监督分类模型等。相关工作多次发表在重要的国际会议和期刊上，研究成果已应用于秦岭卫星遥感自然资源监测，为新型的遥感图像分析和解译方法提供了有力的技术支撑。