Resampling Augmentation for Time Series Contrastive Learning: Application to Remote Sensing

TL;DR

This paper proposes a simple resampling-based augmentation strategy for contrastive learning on satellite image time series, improving agricultural classification performance without complex encodings.

Contribution

Introduction of a novel resampling augmentation method that enhances contrastive learning for remote sensing time series, outperforming existing augmentation techniques.

Findings

Outperforms jittering, resizing, and masking augmentations.

Achieves state-of-the-art results on S2-Agri100 dataset.

Effective without using spatial or temporal encodings.

Abstract

Given the abundance of unlabeled Satellite Image Time Series (SITS) and the scarcity of labeled data, contrastive self-supervised pretraining emerges as a natural tool to leverage this vast quantity of unlabeled data. However, designing effective data augmentations for contrastive learning remains challenging for time series. We introduce a novel resampling-based augmentation strategy that generates positive pairs by upsampling time series and extracting disjoint subsequences while preserving temporal coverage. We validate our approach on multiple agricultural classification benchmarks using Sentinel-2 imagery, showing that it outperforms common alternatives such as jittering, resizing, and masking. Further, we achieve state-of-the-art performance on the S2-Agri100 dataset without employing spatial information or temporal encodings, surpassing more complex masked-based SSL frameworks. Our method offers a simple, yet effective, contrastive learning augmentation for remote sensing time series.