基于U-net的“高分五号”卫星高光谱图像土地类型分类

Land Classification of GF-5 Satellite Hyperspectral Images Using U-net Model

“高分五号”卫星是世界首颗实现对大气和陆地综合观测的全谱段高光谱卫星,对于土地利用类型分类具有重要的应用价值,如何利用深度学习技术开展高光谱图像分类是当前研究的热点问题。深度学习中的语义分割方法在地面场景的图像中已经获得较好的应用,但是对于高光谱遥感图像的精度和适用性较差,无法准确获得精确的分类结果。文章采用U-net模型开展高光谱土地利用类型分类研究,首先基于“高分五号”卫星高光谱数据,构建样本数据集,然后训练分类模型,进行土地利用类型分类,探讨语义分割方法在高分五号高光谱数据上的应用能力。结果表明,采用深度学习中的语义分割方法能够有效提高精度水平,U-net模型的整体分类精度为0.9357,Kappa系数达到0.92,均高于SVM方法和CNN方法。采用深度学习中的语义分割方法,可以为“高分五号”高光谱数据的土地利用分类提供技术支撑,有效提升“高分五号”卫星的应用能力。

GF-5 satellite is the world's first full-spectrum hyper-spectral imagery satellite to observe the Earth’s atmosphere and surface.It is important for the classification of land use types.How to use deep learning technology to carry out hyperspectral image classification is a hot issue in current research.Semantic segmentation method in depth learning has been well applied in the image of ground scene,but the accuracy and applicability of hyperspectral remote sensing images are relatively poor,and the accurate classification results are difficult to be obtained.In this paper,the U-net model is used to study the classification of hyperspectral land use types.Firstly,based on the hyperspectral data of GF-5 satellite,the sample data set is constructed,then the classification model is trained,the land use type classification is carried out,and the application ability of semantic segmentation method on hyperspectral data of GF-5 satellite is discussed.The results show that the semantic segmentation method in deep learning can effectively improve the accuracy level.The overall classification accuracy of the U-net model is 0.9357,and the Kappa coefficient is 0.92,which is higher than the SVM method and CNN method.Using the semantic segmentation method in deep learning,it can provide technical support for land use classification of GF-5 satellite hyperspectral data,and effectively improve the application ability of GF-5 satellite.