高光谱遥感图像分类训练样本问题研究

Hyperspectral remote sensing can well capture both the spatial structural information and rich spectral information of the land covers. More importantly, the spectral information makes an important role in discriminating the materials of different land covers. In recent years, hyperspectral remote sensing has been widely used in many real applications such as accurate agriculture, geological exploration, environmental surveillance, and military reconnaissance. This project aims at proposing novel hyperspectral image classification methods to address the major challenges of hyperspectral image classification, i.e., the small sample size, unreliable sample, and multi-label sample problems as follows: (1) Building the experimental platform for hyperspectral image classification and investigating the mechanism of the sample problems in hyperspectral image classification; (2) Constructing novel collaborative learning based spectral-spatial classification frameworks so as to break the performance limitation of existing small sample size hyperspectral image classification methods; (3) Researching feature and prior knowledge driven noisy label detection and correction method. By this way, the influence of mislabeled samples to the performance of classification can be reduced. (4) Constructing novel multi-label deep learning networks to classify the land covers, so as to reduce human efforts in labeling training samples and improve the classification accuracy. The proposed methods will be applied in several real applications such as detection of red tides and oil spills in ocean, and wetlands surveillance research, so as to provide theoretical and technical supports for ocean pollution reduction and coastal wetland protection.

高光谱遥感能同时获取地物空间结构信息与反映地物材质属性的精细光谱信息，具有“图谱合一”的特点，被广泛应用于精准农业、地质勘探、环境监测、军事侦察等重要领域。本项目拟深入研究高光谱遥感图像监督分类面临的小样本、不可靠样本、多标签样本等关键问题，并提出针对性的高光谱图像分类新方法。研究内容包括：（1）构建高光谱图像分类实验平台，深入研究各类训练样本问题的产生机理；（2）提出基于协同学习的高光谱图像分类方法，突破现有小样本图像分类方法的性能瓶颈；（3）提出特征与先验知识共同驱动的不可靠样本检测与校正方法，降低不可靠样本对分类性能的影响；（4）提出针对多标签样本的高光谱图像深度学习分类方法，降低人工标记成本，提升高光谱图像分类精度。本项目研究成果将应用于海洋赤潮、溢油检测以及海岸带湿地植被监测等领域，为我国海洋排污减灾与海岸带湿地保护提供理论与技术支撑。

高光谱图像高效智能处理是遥感领域的前沿挑战难题。项目围绕高光谱图像分类样本问题开展了深入研究，成功揭示了多类训练样本问题的产生机理；针对在成像过程中数据受到多种退化因素干扰的问题，提出了基于退化模型引导的高光谱图像复原技术，解决了图像受干扰时分类精度低的问题；针对目标与背景的区分性较低的问题，提出了基于光谱反射特性引导的深度网络的对比度增强方法与RGB图像引导下的高光谱图像分辨率增强方法，提升了分类精度。针对小样本问题，提出了多种多尺度空谱特征提取方法，设计了结构引导的单样本高光谱图像分类框架，实现了小样本情形下地物的高精度识别。针对不可靠样本问题，提出了基于空间密度峰值聚类的不可靠样本检测与校正方法。针对存在噪声样本的问题，提出了图像噪声不敏感的结构特征提取技术。此外，设计了与抗噪声深度邻域嵌入损失函数，利用图像结构来降低潜在噪声图像对网络训练的贡献权重，实现了更鲁棒的高光谱图像分类。针对样本标记耗时耗力的问题，构建了高光谱图像的高效标注策略（多标签样本）与相应的训练模型，并在此基础上提出了基于超像素分割和递归滤波的多标签样本分类方法。项目成果成功应用于湿地资源调查、海洋灾害监测等重要领域，取得了很好的社会效益，获2021年度湖南省自然科学一等奖1项。