MSAR数据集

数据主编：陈杰 黄志祥 夏润繁 大规模多类SAR目标检测数据集-1.0（MSAR-1.0）共包括28449张检测切片，采用海丝一号卫星和高分三号卫星数据。 MSAR-1.0数据集极化方式包括HH、HV、VH和VV。该数据集场景包括机场、港口、近岸、岛屿、远海、城区等；类型包括飞机、油罐、桥梁和船只四类目标，由1851架桥梁，39858条船只，12319个油罐和6368架飞机组成。图1是MSAR-1.0数据集的部分切片样例。 图1. MSAR-1.0数据集样例. (a)和(b)是油罐；(c)和(d)是船只；(e)和(f)是桥梁；(g)和(h)是飞机 MSAR-1.0数据集切片尺寸为256×256像素，部分桥梁切片为2048×2048像素，格式为三通道灰度图像，24位深JPG。标注格式为XML格式，记录目标类型和位置信息，其中位置信息由Xmin、Xmax、Ymin和Ymax组成。MSAR-1.0数据集切片标签文件示例见下面数据集使用说明中图2、3。符合Yolo系列、PolarMask、SSD和Faster-RCNN等主流检测网络的格式要求。 为满足更多科研需求，MSAR-1.0数据集提供了多幅海丝一号卫星拍摄的大场景图像，类型包括飞机、油罐、桥梁、船只、机场跑道等。标注格式为XML格式，记录目标类型和位置信息，其中位置信息由Xmin、Xmax、Ymin和Ymax组成。MSAR-1.0数据集大场景图像标签文件示例见数据集使用说明中图4、5。 数据集使用说明见“大规模多类SAR目标检测数据集-1.0使用说明”。 本期数据引用格式如下： [1] 陈杰, 黄志祥, 夏润繁，邬伯才，盛磊，孙龙，姚佰栋. 大规模多类SAR目标检测数据集-1.0[OL]. 雷达学报, 2022. https://radars.ac.cn/web/data/getData?dataType=MSAR. Jie Chen, Zhixiang Huang, Runfan Xia, Bocai Wu, Lei Sheng, Long Sun, and Baidong Yao. Large-scale multi-class SAR image target detection dataset-1.0[OL]. Journal of Radars, 2022. https://radars.ac.cn/web/data/getData?dataType=MSAR. [2] Xia, R.; Chen, J.; Huang, Z.; Wan, H.; Wu, B.; Sun, L.; Yao, B.; Xiang, H.; Xing, M. CRTransSar: A Visual Transformer Based on Contextual Joint Representation Learning for SAR Ship Detection. Remote Sensing. 2022, 14, 1488

Data Editors: Jie Chen, Zhixiang Huang, Runfan Xia

The Large-scale Multi-class SAR Target Detection Dataset-1.0 (MSAR-1.0) contains a total of 28,449 detection patches, which are derived from data collected by the HaiSi-1 and Gaofen-3 (GF-3) satellites.

The polarization modes of the MSAR-1.0 dataset include HH, HV, VH, and VV.

The dataset covers scenarios such as airports, ports, nearshore areas, islands, open seas, and urban areas, with four target categories: aircraft, oil tanks, bridges, and ships. The dataset comprises 1,851 bridges, 39,858 ships, 12,319 oil tanks, and 6,368 aircraft.

Figure 1 shows partial sample patches of the MSAR-1.0 dataset.
Figure 1. Sample patches of the MSAR-1.0 dataset. (a) and (b) are oil tanks; (c) and (d) are ships; (e) and (f) are bridges; (g) and (h) are aircraft.

The patch size of the MSAR-1.0 dataset is 256×256 pixels, with some bridge patches being 2048×2048 pixels. All patches are three-channel grayscale images in 24-bit JPG format.

The annotation format is XML, which records the target category and location information, where the location information is defined by Xmin, Xmax, Ymin, and Ymax.

Examples of the patch annotation files for the MSAR-1.0 dataset are shown in Figures 2 and 3 of the Dataset User Guide. This dataset complies with the format requirements of mainstream detection networks such as YOLO series, PolarMask, SSD, and Faster-RCNN.

To meet additional research requirements, the MSAR-1.0 dataset also provides multiple large-scene images captured by the HaiSi-1 satellite, covering targets such as aircraft, oil tanks, bridges, ships, airport runways, etc. The annotation format is XML, which records the target category and location information defined by Xmin, Xmax, Ymin, and Ymax.

Examples of the large-scene image annotation files for the MSAR-1.0 dataset are shown in Figures 4 and 5 of the Dataset User Guide.

The Dataset User Guide is titled *Large-scale Multi-class SAR Target Detection Dataset-1.0 User Guide*.

The citation formats for this dataset are as follows:
[1] Chen J, Huang Z, Xia R, Wu B, Sheng L, Sun L, Yao B. Large-scale multi-class SAR image target detection dataset-1.0[OL]. Journal of Radars, 2022. https://radars.ac.cn/web/data/getData?dataType=MSAR.
[2] Xia R, Chen J, Huang Z, Wan H, Wu B, Sun L, Yao B, Xiang H, Xing M. CRTransSar: A Visual Transformer Based on Contextual Joint Representation Learning for SAR Ship Detection. Remote Sensing, 2022, 14, 1488.