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

TL;DR

This paper introduces SSL4EO-S12, a large-scale, multi-modal, multi-temporal satellite dataset for self-supervised learning in Earth observation, demonstrating its effectiveness across various pre-training methods and surpassing existing datasets.

Contribution

The authors present a new extensive unlabeled satellite dataset, SSL4EO-S12, and validate its utility for self-supervised pre-training in Earth observation tasks, achieving competitive or superior results.

Findings

Pre-training on SSL4EO-S12 improves downstream accuracy.

Models pre-trained on SSL4EO-S12 outperform those on existing datasets.

The dataset supports multiple self-supervised learning methods.

Abstract

Self-supervised pre-training bears potential to generate expressive representations without human annotation. Most pre-training in Earth observation (EO) are based on ImageNet or medium-size, labeled remote sensing (RS) datasets. We share an unlabeled RS 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 from the ESA Sentinel-1 \& -2 satellite missions. For EO applications we demonstrate SSL4EO-S12 to succeed in self-supervised pre-training for a set of methods: MoCo-v2, DINO, MAE, and data2vec. Resulting models yield downstream performance close to, or surpassing accuracy measures of supervised learning. In addition, pre-training on SSL4EO-S12 excels compared to existing datasets. We make openly available the dataset, related source code, and pre-trained models at https://github.com/zhu-xlab/SSL4EO-S12.