BigEarthNet: A Large-Scale Benchmark Archive For Remote Sensing Image Understanding
BigEarthNet, a large-scale multi-label Sentinel-2 dataset with 590,326 images, outperforms ImageNet pre-trained models in land-cover scene classification.
Key Findings
Methodology
This study constructed BigEarthNet, comprising 590,326 multi-label Sentinel-2 image patches, annotated using the 2018 CORINE Land Cover database. A shallow CNN architecture was trained from scratch and compared with transfer learning using Inception-v2 pre-trained on ImageNet. The images included RGB and full spectral bands, processed through interpolation and cloud masking. Data was split into training, validation, and test sets, and model performance was evaluated via multi-label F1-score, precision, and recall.
Key Results
- Models trained from scratch on BigEarthNet achieved an F1-score of 0.7098, surpassing the 0.4988 score of the ImageNet-based transfer model by approximately 42%. Incorporating all spectral bands further improved F1 to 0.7384.
- Results demonstrate that training directly on large-scale, multi-label remote sensing data yields superior performance compared to transfer learning, emphasizing the importance of dataset scale.
- Multi-label annotations significantly enhance the model's ability to recognize complex scenes, with consistent performance across different seasons and geographic regions.
Significance
This work addresses the critical bottleneck of data scarcity in remote sensing deep learning, providing a comprehensive, annotated dataset that enables more accurate and robust land-cover classification. It bridges the gap between computer vision datasets and remote sensing needs, fostering advancements in environmental monitoring, urban planning, and disaster management. The availability of BigEarthNet paves the way for more sophisticated models and operational applications, marking a milestone in remote sensing AI research.
Technical Contribution
The paper introduces the first large-scale, multi-label Sentinel-2 dataset, leveraging CORINE Land Cover annotations for detailed land-cover categorization. It demonstrates that training shallow CNNs from scratch on this dataset outperforms transfer learning from ImageNet, highlighting the importance of domain-specific large datasets. The preprocessing pipeline includes spectral interpolation and cloud masking, ensuring high-quality inputs. The study also provides a comparative analysis of training strategies, offering new insights into remote sensing deep learning workflows.
Novelty
This is the first large-scale, multi-label Sentinel-2 dataset aligned with CORINE land cover classes, significantly larger and more detailed than existing datasets. Unlike prior works limited to single-label or small datasets, BigEarthNet enables training deep models directly on remote sensing data, emphasizing the importance of dataset scale and multi-label annotations for real-world scene complexity.
Limitations
- Despite its size, the dataset still contains cloud-covered and seasonal snow regions, which may affect model generalization. Further refinement in cloud masking is needed.
- 浅层CNN模型在复杂场景和高分辨率细节捕获方面仍有限，未来应探索更深层网络架构。
- 目前未充分利用时间序列信息，结合多时相数据可能进一步提升模型性能。
Future Work
未来将持续扩展BigEarthNet，增加多时相、多源、多尺度数据，丰富标签体系。计划引入深层网络结构和多任务学习框架，提升模型泛化能力。结合地理信息系统（GIS）数据实现空间关系建模，推动遥感智能应用的实用化。
AI Executive Summary
Remote sensing has long been constrained by limited data availability and simplistic labels, hindering the application of deep learning techniques. Existing datasets like UC Merced and WHU-RS19 are small and single-labeled, insufficient for training modern neural networks. To address this, the authors introduce BigEarthNet—a comprehensive dataset with 590,326 multi-label Sentinel-2 image patches, annotated using the 2018 CORINE Land Cover database.
This dataset covers ten European countries, with images sized at 120×120, 60×60, and 20×20 pixels, capturing diverse land cover types. The authors designed a shallow CNN with three convolutional layers, trained from scratch, and compared it with transfer learning using Inception-v2 pre-trained on ImageNet. Results show the from-scratch model achieved an F1-score of 0.7098, significantly outperforming the transfer model (F1=0.4988), demonstrating the power of large-scale, domain-specific data.
The experiments confirmed that training directly on BigEarthNet yields superior accuracy, especially when using all spectral bands. The multi-label annotations enable the model to recognize complex scenes, making it highly applicable for land use, environmental monitoring, and disaster response. The authors plan to expand the dataset further, incorporating multi-temporal and multi-source data, and exploring deeper neural architectures. This work sets a new benchmark for remote sensing AI, promising to accelerate research and operational deployment in Earth observation.
Deep Analysis
Background
遥感技术经过多年的发展，已广泛应用于土地利用、环境监测、灾害评估等领域。早期研究多依赖手工特征或浅层模型，受限于数据规模和标注信息。近年来，深度学习如卷积神经网络（CNN）在图像识别中取得突破，但遥感领域缺乏大规模、多标签、多光谱的公开数据集，限制了模型性能提升。现有数据集如UC Merced、WHU-RS19等，规模较小，标签单一，难以满足深度模型训练需求。为解决这一问题，学界开始探索大规模遥感数据集，但多标签、多源、多尺度的集成仍是挑战。
Core Problem
当前遥感深度学习面临数据不足、标签单一、模型泛化差等核心问题。缺少大规模、多标签、多光谱的公开数据集，导致模型难以捕获复杂场景的多样性。迁移学习虽然缓解部分问题，但受限于源数据与遥感数据的差异，效果有限。如何构建规模大、标签丰富、质量高的遥感数据集，成为制约行业发展的瓶颈。此外，现有模型在复杂环境、多类别、多尺度场景中的表现仍不理想，亟需更丰富的数据和更强的模型。
Innovation
本研究的创新点在于：1）首次构建了包含590,326个多标签Sentinel-2图像块的BigEarthNet，结合CORINE 2018土地覆盖数据库实现多类别标注，突破了以往规模小、单标签的限制；2）采用多光谱插值和云遮挡剔除技术，确保数据质量；3）验证浅层CNN在大规模多标签数据上的优越性，超越迁移自ImageNet的深层模型。该数据集的多标签设计更贴近实际应用场景，提升模型识别复杂场景的能力。
Methodology
- �� 数据采集：选择2017年6月至2018年5月欧洲10国的Sentinel-2影像，云量<1%，经过大气校正。• 图像预处理：剔除云遮挡区域，插值补全20m和60m波段，统一像素大小。• 图像切片：将大影像切割成120×120、60×60和20×20像素块，形成590,326个样本。• 标注：利用CORINE 2018数据库，为每个图像块赋予1-12个土地覆盖类别的多标签。• 模型训练：设计浅层CNN（3卷积层+全连接层），用随机参数初始化，训练目标为多标签交叉熵损失。• 评估：划分训练、验证、测试集，计算F1-score、精确率、召回率。• 比较：从零训练模型与迁移学习模型（Inception-v2预训练）性能对比。
Experiments
采用随机划分的训练集（60%）、验证集（20%）和测试集（20%），训练100轮，使用SGD优化。模型输入包括RGB和全光谱数据，后者通过插值处理。性能指标为多标签F1-score、精确率和召回率。对比实验显示，从零训练的浅层CNN在所有指标上优于迁移模型，验证了大规模多标签数据的有效性。还分析了不同光谱信息对性能的影响，结果表明全光谱模型表现更佳。
Results
实验结果显示，浅层CNN在BigEarthNet上训练，F1-score达0.7098，优于迁移模型（F1=0.4988），提升约42%。考虑全部光谱后，性能进一步提升，F1达0.7384。多标签标注显著增强模型识别复杂场景的能力，尤其在土地覆盖多样性区域表现优越。模型在不同季节和地区均保持稳定，验证了数据集的代表性和模型的泛化能力。这些结果证明了BigEarthNet在遥感深度学习中的应用潜力。
Applications
该数据集可广泛应用于土地利用分类、环境监测、灾害评估等场景，支持深度学习模型的训练和优化。企业和科研机构可以利用BigEarthNet开发自动化土地覆盖识别、变化检测等系统，提升处理效率和准确性。未来，结合多源数据和深层网络，有望实现更高精度和更复杂场景的智能分析，推动遥感行业的智能化转型。
Limitations & Outlook
尽管规模庞大，数据集仍存在云遮挡、季节变化等影响，部分图像质量不足，可能影响模型泛化。浅层CNN模型在复杂场景中的表现有限，深层网络需进一步探索。当前未充分利用时空信息，未来应结合多时相、多源数据提升模型能力。此外，数据标注依赖自动化和人工校验，仍存在误差风险。
Plain Language Accessible to non-experts
想象你在一个大型工厂里，工厂每天生产各种不同的产品。每个产品都需要经过不同的检测步骤，确保质量。以前，检测人员只能用肉眼判断产品是否合格，效率低、容易出错。现在，工厂引入了一台智能检测机器，它可以同时观察产品的多个方面，比如颜色、形状、大小等，还能识别出产品属于哪个类别。这个机器就像BigEarthNet一样，是一个庞大的“检测库”，里面存放了成千上万的产品图片，每个图片都标注了多个特征。通过训练这台机器，让它学会识别各种复杂的产品特征，工厂的检测效率大大提高，产品质量也更有保障。这就像遥感中的图像分类，BigEarthNet帮助机器更聪明地理解地球表面各种土地类型。
ELI14 Explained like you're 14
想象你在一个超级大的学校里，有很多不同的班级和学生。以前，老师只能用简单的方式知道哪个学生属于哪个班，比如只看名字，不能知道他们的兴趣、爱好或特长。现在，学校引入了一个智能系统，它可以同时知道每个学生的多项信息，比如喜欢的运动、擅长的科目、兴趣爱好等等。这个系统学习了很多学生的资料，变得非常聪明，能帮老师更好地了解每个学生。BigEarthNet就像这个智能系统，它收集了成千上万的地球图片，每张图片都标注了多个类别，比如森林、城市、湖泊等。通过让电脑学习这些图片，电脑变得更聪明，能更准确地识别地球上的不同土地类型。这就像学校里的老师变得更懂学生一样，BigEarthNet让机器更懂地球。
Glossary
Sentinel-2（哨兵-2）
欧洲空间局的多光谱遥感卫星，提供高分辨率多光谱图像，广泛用于土地覆盖和环境监测。
本文使用Sentinel-2影像作为数据源，进行大规模场景分类。
多标签（Multi-label）
每个图像可以同时属于多个类别，而非单一类别，反映复杂场景的多样性。
BigEarthNet中的图像都带有多标签，增强模型识别能力。
CORINE Land Cover（CORINE土地覆盖）
欧洲环境局发布的土地覆盖数据库，提供详细的土地类型分类信息。
用于为BigEarthNet图像标注多类别标签。
卷积神经网络（CNN）
一种深度学习模型，擅长处理图像数据，通过卷积层提取空间特征。
本文采用浅层CNN进行场景分类。
F1-score（F1分数）
精确率和召回率的调和平均，用于衡量多标签分类模型性能。
实验中用以评估模型的整体表现。
Open Questions Unanswered questions from this research
- 1 如何进一步利用时空信息和多源数据提升遥感图像的多标签分类性能仍是未解难题。
- 2 深层网络在大规模多标签遥感数据上的训练策略和优化方法有待深入研究。
- 3 多标签标注的误差对模型性能的影响及其校正机制尚不充分理解。
Applications
Immediate Applications
土地利用监测
利用BigEarthNet训练模型，实现自动识别不同土地类型，支持城市规划和环境保护。
灾害应急响应
快速分析受灾区域的土地变化，辅助决策，提升应急效率。
Long-term Vision
智能地理信息系统
结合大数据和深度学习，构建全自动化、实时更新的土地覆盖监测平台，推动智慧城市和精准农业发展。
Abstract
This paper presents the BigEarthNet that is a new large-scale multi-label Sentinel-2 benchmark archive. The BigEarthNet consists of 590,326 Sentinel-2 image patches, each of which is a section of i) 120x120 pixels for 10m bands; ii) 60x60 pixels for 20m bands; and iii) 20x20 pixels for 60m bands. Unlike most of the existing archives, each image patch is annotated by multiple land-cover classes (i.e., multi-labels) that are provided from the CORINE Land Cover database of the year 2018 (CLC 2018). The BigEarthNet is significantly larger than the existing archives in remote sensing (RS) and thus is much more convenient to be used as a training source in the context of deep learning. This paper first addresses the limitations of the existing archives and then describes the properties of the BigEarthNet. Experimental results obtained in the framework of RS image scene classification problems show that a shallow Convolutional Neural Network (CNN) architecture trained on the BigEarthNet provides much higher accuracy compared to a state-of-the-art CNN model pre-trained on the ImageNet (which is a very popular large-scale benchmark archive in computer vision). The BigEarthNet opens up promising directions to advance operational RS applications and research in massive Sentinel-2 image archives.