天宫一号高光谱成像仪具有空间分辨率高、光谱分辨率高、图谱合一等特性，在中国航天高光谱领域具有里程碑的意义。针对一般遥感场景分类数据集尺度单一、光谱分辨率较低等问题，本文提出基于天宫一号的多谱段、高空间分辨率、多时相高光谱遥感场景分类数据集(TG1HRSSC)。利用天宫一号高光谱成像仪获取的高质量数据，经过辐射校正、几何校正、空间裁剪、波段筛选、数据质量分析与控制等，制作了一批通用的航天高光谱遥感场景分类数据集，通过载人航天空间应用数据推广服务平台（ http://www.msadc.cn [2019-09-10]）进行分发和共享。该数据集包括天宫一号高光谱成像仪获取的城镇、农田、林地、养殖塘、荒漠、湖泊、河流、港口、机场等9个典型地物场景的204个高光谱影像数据，其中5 m分辨率全色谱段1个波段、10 m分辨率可见近红外谱段54个有效波段以及20 m分辨率短波红外谱段52个有效波段。研究利用AlexNet、VGG-VD-16、GoogLeNet等深度学习算法网络对构建的数据集进行场景分类的试验，结果表明该数据集的场景分类应用实现较好效果。由于该数据集具备高分辨、高光谱等特征优势，未来在语义理解、多目标检测等方面有着广泛的应用价值。

Remote sensing image scene classification is an important means of remote sensing image interpretation, which has important application value in land and resources investigation, ecological environment monitoring, disaster assessment, target interpretation and so on. In recent years, deep learning has become a research hotspot in the field of remote sensing scene classification, and data set is the basis for its development. Most of the existing remote sensing scene classification datasets are true color images with single scale and less spectral information. And other hyperspectral data sets have relatively small data coverage. The data of Tiangong-1 Hyperspectral Imager has the characteristics of high spatial resolution, high hyperspectral resolution and wide coverage. It can be used for comprehensive feature extraction and analysis of spectral spatial information, which can provide more abundant data sources for remote sensing image classification application research, and make up for the deficiency of spectral information and limited application of common remote sensing scene classification data sets. In this study, based on the high-quality data acquired by Tiangong-1 Hyperspectral Imager, Tiangong-1 Hyperspectral Remote Sensing Scene Classification data set (TG1HRSSC) is produced through radiation correction, geometric correction, spatial clipping, band screening, and data quality analysis and control. The dataset, which contains the 204 hyperspectral multiresolution image data of nine typical scenes (e.g., city, farmland, forest, pond culture, desert, lake, river and airport), is published and shared in the Space Application Data Promoting Service Platform for China Manned Space Engineering ( http://www.msadc.cn [2019-09-10]). The dataset includes one band of 5 m resolution full spectrum, 54 bands of 10 m resolution visible and near-infrared spectrum, and 52 bands of 20 m resolution short-wave infrared spectrum. In addition, this paper describes and analyzes the data set from four aspects: scene distribution, time distribution, spectral distribution and scale distribution. In order to test the application effect of data classification, three classical convolution neural networks (VGG-VD-16, AlexNet and GoogleLeNet) are selected to train the data sets by transfer learning. The overall classification accuracy is 91.52 ± 0.60, 90.47 ± 0.23 and 89.12 ± 0.34, respectively. Results show that the scene classification of the dataset is effective. In following research, the network model can be designed to make full use of the multi-spectral characteristics of the data to achieve more accurate scene classification, and to improve the generalization ability of existing models by using the characteristics of multi-scale data. The data set (TG1HRSSC) has the advantages of hyperspectral, high spatial resolution and multi-scale. The abundant spectral information and fine spatial information provide data support for the research of target recognition of fine ground objects, remote sensing scene classification, remote sensing semantic understanding and other applications, which has unique value and application prospects.