本数据集为“MLRSNet”的新型大规模高分辨率多标签遥感数据集，用于语义场景理解。该数据集包含109,161张高分辨率遥感图像，这些图像被标注为46个类别，每个类别的样本图像数量在1500到3000之间。这些图像的固定尺寸为 256×256像素，且具有不同的像素分辨率。此外，数据集中的每张图像都标注了60个预定义类别标签中的多个，每个图像关联的标签数量在1到13之间。此外，详细阐述了MLRSNet数据集的构建流程，并评估了多种基于多标签的深度学习方法在图像分类和图像检索任务中的性能。实验结果表明，基于多标签的深度学习方法在图像分类和图像检索任务中能取得更优性能。该数据集是目前最大的高分辨率多标签遥感数据集，包含最丰富的多标签信息。且该数据集具有较高的类内多样性，可为语义场景理解领域中众多方法的评估与发展提供更优质的数据资源。", "ds_source": "
MLRSNet 由来自世界各地的 109,161 张标记的 RGB 图像组成，分为 46 大类：飞机、机场、裸地、棒球场、篮球场、海滩、桥梁、丛林、云、商业区、密集住宅区、沙漠、侵蚀农田、农田、森林、高速公路、高尔夫球场、地面田径场。港口和港口、工业区、十字路口、岛屿、湖泊、草地、移动房屋公园、 山、立交桥、公园、停车场、公园大道、铁路、火车站、河流、环形交叉路口、船坞、雪山、稀疏的住宅区、体育场、储罐、游泳池、网球场、露台、输电塔、蔬菜温室、湿地和风力涡轮机。样本图像的数量因不同的大类而变化很大，从1500张到3000张不等。此外，数据集中的每张图像都分配了 60 个预定义类标签中的几个，并且与每个图像关联的标签数量在 1 到 13 之间变化。", "ds_process_way": "
MLRSNet的建设主要由场景样本采集、数据库质量控制和数据库样本多样性改进三个过程组成。", "ds_quality": "
相比之下，与现有的遥感图像数据集相比，MLRSNet具有以下显著特点：
1、层次结构：MLRSNet包含3个一级类别，如土地利用与土地覆盖（例如商业区、农田、森林、工业区、山地）、自然物体与地貌（例如海滩、云层、岛屿、湖泊、河流、灌木丛），以及人工物体与地貌（例如飞机、机场、桥梁、高速公路、立交桥）， 46 个二级类别和 60 个三级标签。
2、多标签：MLRSNet数据集中的每张图像都对应一个或多个标签，因为遥感图像通常包含多种不相互排斥的物体类别。多项实验表明，在图像分类或图像检索任务中，多标签数据集往往能取得比单标签数据集更优异的性能。
3、大规模：MLRSNet包含大量高分辨率多标签遥感场景图像。该数据集包含109,161张高分辨率遥感图像，被标注为46个类别，每个类别的样本图像数量在1500至3000之间，均大于其他大多数列出的数据集。MLRSNet是一个为场景图像识别收集的大规模高分辨率遥感数据集，可覆盖更广泛的卫星或航空图像范围。它旨在作为替代方案，推动场景图像识别方法的发展，特别是需要大量标注训练数据的深度学习方法。
4、多样性：为提升数据集的泛化能力，我们尝试根据地理分布、季节分布、天气条件、视角、采集时间及图像分辨率等维度对 MLRSNet 进行特征描述，即在空间分辨率、视角、物体姿态、光照、背景以及遮挡等方面存在显著差异。
", "ds_acq_start_time": null, "ds_acq_end_time": null, "ds_acq_place": "", "ds_acq_lon_east": null, "ds_acq_lat_south": null, "ds_acq_lon_west": null, "ds_acq_lat_north": null, "ds_acq_alt_low": null, "ds_acq_alt_high": null, "ds_share_type": "login-access", "ds_total_size": 1314947133, "ds_files_count": 2, "ds_format": "jpg,csv", "ds_space_res": "10-0.1米", "ds_time_res": "", "ds_coordinate": "无", "ds_projection": "", "ds_thumbnail": "0165a0f2-6ded-41f4-a94b-6d0aceb3d199.png", "ds_thumb_from": 2, "ds_ref_way": "", "paper_ref_way": "", "ds_ref_instruction": "", "ds_from_station": null, "organization_id": "52b7b79b-860c-49a5-9083-9a70cf8bed5a", "ds_serv_man": "敏玉芳", "ds_serv_phone": "09314967596", "ds_serv_mail": "ncdc@lzb.ac.cn", "doi_value": "", "subject_codes": [ "170.45" ], "quality_level": 3, "publish_time": "2025-07-30 15:15:41", "last_updated": "2026-01-14 10:57:02", "protected": false, "protected_to": null, "lang": "zh", "cstr": "11738.11.NCDC.NIEER.DB6932.2025", "i18n": { "en": { "title": "MLRSNet: A Multi Label High Spatial Resolution Remote Sensing Dataset for Semantic Scene Understanding", "ds_format": "jpg,csv", "ds_source": "MLRSNet consists of 109161 labeled RGB images from around the world, divided into 46 categories: airplanes, airports, bare ground, baseball fields, basketball courts, beaches, bridges, jungles, clouds, commercial areas, densely populated residential areas, deserts, eroded farmland, farmland, forests, highways, golf courses, and ground athletics fields. Ports and harbors, industrial areas, intersections, islands, lakes, grasslands, mobile home parks, mountains, overpasses, parks, parking lots, park avenues, railways, train stations, rivers, roundabouts, shipyards, snow capped mountains, sparse residential areas, sports fields, storage tanks, swimming pools, tennis courts, terraces, transmission towers, vegetable greenhouses, wetlands, and wind turbines. The number of sample images varies greatly depending on different categories, ranging from 1500 to 3000. In addition, each image in the dataset is assigned several of the 60 predefined class labels, and the number of labels associated with each image varies from 1 to 13.", "ds_quality": "
Compared with existing remote sensing image datasets, MLRSNet has the following significant characteristics:
1. Hierarchical structure: MLRSNet consists of three primary categories, such as land use and land cover (e.g. commercial areas, farmland, forests, industrial areas, mountains), natural objects and landforms (e.g. beaches, clouds, islands, lakes, rivers, shrubs), and artificial objects and landforms (e.g. airplanes, airports, bridges, highways, overpasses), with 46 secondary categories and 60 tertiary labels.
2. Multi label: Each image in the MLRSNet dataset corresponds to one or more labels, as remote sensing images typically contain multiple non mutually exclusive object categories. Multiple experiments have shown that in image classification or retrieval tasks, multi label datasets often achieve better performance than single label datasets.
3. Large scale: MLRSNet contains a large number of high-resolution, multi label remote sensing scene images. This dataset contains 109161 high-resolution remote sensing images, labeled as 46 categories, with sample images ranging from 1500 to 3000 for each category, all larger than most other listed datasets. MLRSNet is a large-scale high-resolution remote sensing dataset collected for scene image recognition, which can cover a wider range of satellite or aerial images. It aims to serve as an alternative solution to promote the development of scene image recognition methods, especially deep learning methods that require a large amount of annotated training data.
4. Diversity: To enhance the generalization ability of the dataset, we attempted to describe the features of MLRSNet based on dimensions such as geographic distribution, seasonal distribution, weather conditions, perspective, acquisition time, and image resolution. Specifically, there were significant differences in spatial resolution, perspective, object pose, lighting, background, and occlusion.
", "ds_ref_way": "", "ds_abstract": "This dataset is a new large-scale high-resolution multi label remote sensing dataset developed by \"MLRSNet\" for semantic scene understanding. This dataset contains 109161 high-resolution remote sensing images labeled as 46 categories, with sample images ranging from 1500 to 3000 for each category. These images have a fixed size of 256 × 256 pixels and different pixel resolutions. In addition, each image in the dataset is labeled with multiple of the 60 predefined category labels, with each image associated with a label count ranging from 1 to 13. In addition, the construction process of the MLRSNet dataset was elaborated in detail, and the performance of various multi label based deep learning methods in image classification and image retrieval tasks was evaluated. The experimental results show that multi label based deep learning methods can achieve better performance in image classification and image retrieval tasks. This dataset is currently the largest high-resolution multi label remote sensing dataset, containing the richest multi label information. And this dataset has high intra class diversity, which can provide better data resources for the evaluation and development of numerous methods in the field of semantic scene understanding.
", "ds_time_res": "", "ds_acq_place": "Global", "ds_space_res": "10-0.1米", "ds_projection": "", "ds_process_way": "The construction of MLRSNet mainly consists of three processes: scene sample collection, database quality control, and improvement of database sample diversity.", "ds_ref_instruction": "" } }, "submit_center_id": "ncdc", "data_level": 0, "license_type": "https://creativecommons.org/licenses/by/4.0/", "doi_reg_from": "reg_outside", "cstr_reg_from": "reg_outside", "doi_not_reg_reason": null, "cstr_not_reg_reason": null, "is_paper_in_submitting": false, "ds_topic_tags": [ "多标签图像数据集", "语义场景理解", "卷积神经网络（CNN）", "图像分类", "图像检索" ], "ds_subject_tags": [ "地理学" ], "ds_class_tags": [], "ds_locus_tags": [], "ds_time_tags": [], "ds_contributors": [ { "true_name": "王跃宾", "email": "wangyuebin@cugb.edu.cn", "work_for": "中国地质大学", "country": "中国" } ], "ds_meta_authors": [ { "true_name": "王跃宾", "email": "wangyuebin@cugb.edu.cn", "work_for": "中国地质大学", "country": "中国" } ], "ds_managers": [ { "true_name": "王跃宾", "email": "wangyuebin@cugb.edu.cn", "work_for": "中国地质大学", "country": "中国" } ], "category": "遥感及产品" }