SSL4EO-S12: A Large-Scale Multi-Modal, Multi-Temporal Dataset for Self-Supervised Learning in Earth Observation

Self-supervised pre-training bears potential to generate expressive representations from large-scale Earth observation (EO) data without human annotation. However, most existing pre-training in the field is based on ImageNet or medium-size, labeled remote sensing (RS) datasets. In this paper, we share an unlabeled dataset SSL4EO-S12: Self-Supervised Learning for Earth Observation - Sentinel-1/2, to assemble a large-scale, global, multimodal, and multi-seasonal corpus of satellite imagery. We demonstrate SSL4EO-S12 to succeed in self-supervised pre-training for a set of representative methods: MoCo-v2, DINO, MAE and data2vec, and multiple downstream applications including scene classification, semantic segmentation and change detection. Our benchmark results prove the effectiveness of SSL4EO-S12 compared to existing datasets. The dataset, related source code, and pre-trained models are available at https://github.com/zhu-xlab/SSL4EO-S12.

自监督预训练（Self-supervised pre-training）有望在无需人工标注的情况下，从大规模地球观测（Earth Observation, EO）数据中生成具有表达力的表征。然而，该领域现有的大多数预训练方法均基于ImageNet数据集或中等规模的带标注遥感（Remote Sensing, RS）数据集。本文中，我们发布了一个无标注数据集SSL4EO-S12：地球观测自监督学习——哨兵1号/2号（Self-Supervised Learning for Earth Observation - Sentinel-1/2），旨在构建一个大规模、全球覆盖、多模态且多季节的卫星影像语料库。我们验证了SSL4EO-S12在MoCo-v2、DINO、MAE及data2vec等一系列代表性方法的自监督预训练中表现优异，并可支持场景分类、语义分割及变化检测等多种下游应用。基准测试结果表明，与现有数据集相比，SSL4EO-S12具有显著的有效性。该数据集、相关源代码及预训练模型可通过以下链接获取：https://github.com/zhu-xlab/SSL4EO-S12。