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

2025-06-23

AI Summary

Remote sensing faces the dual challenge of scarce labeled satellite image time series (SITS) data and underutilization of abundant unlabeled sequences. To address this, we propose a resampling-based contrastive learning framework specifically designed for SITS: it constructs temporally consistent positive pairs via upsampling and disjoint subsequence extraction, preserving the original temporal coverage without relying on spatial features or explicit temporal encoding. Unlike conventional augmentation strategies—such as jittering, cropping, or masking—our approach avoids distortions that disrupt temporal coherence, thereby significantly improving unsupervised pretraining quality. Evaluated on multiple agricultural classification benchmarks (e.g., S2-Agri100), our method achieves superior performance with a simpler architecture compared to complex mask-based self-supervised approaches, demonstrating both effectiveness and strong generalization across diverse downstream tasks.

Technology Category

• Computer Vision: Remote Sensing / Geospatial AI
• Machine Learning: Semi-Supervised Learning
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

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.

Problem

• Addresses scarcity of labeled Satellite Image Time Series data
• Proposes resampling augmentation for time series contrastive learning
• Improves agricultural classification using Sentinel-2 imagery

Innovation

• Resampling-based augmentation for time series
• Upsampling and disjoint subsequence extraction
• State-of-the-art performance without spatial info