Lecture 2

Self-supervised learning techniques to address small data problems with remote sensing for agriculture

Description: This lecture focuses on Self-Supervised Learning (SSL) techniques and their application to small data problems in remote sensing for agriculture. SSL allows models to use unlabelled data to learn useful representations, making it an ideal solution for data-poor environments. The talk will introduce several SSL models, including SimSiam, SimCLR, MoCo, Barlow Twins, and VICReg, and explain how these models can be applied to remote sensing data using the example of the European Land Use-Land Cover Area Frame Survey (LUCAS), such as crop classification and segmentation. We will also discuss how these SSL methods can be integrated with traditional deep learning models to improve performance when labelled data is limited.

Objective: Develop the ability to apply Self-Supervised Learning techniques, such as SimSiam, SimCLR, MoCo, Barlow Twins, and VICReg, to Remote Sensing data and enhance existing Deep Learning models for small data problems.
Duration: 4 hours

Lecture Flow (content):

  1. Introduction to Self-Supervised Learning and its relevance to small data problems in Remote Sensing (10 min)
  2. Key Self-Supervised Learning models and techniques: (40 min)
    • SimSiam: Simple framework for contrastive learning without negative samples
    • SimCLR: Contrastive Learning for representation learning
    • MoCo: Momentum Contrast for unsupervised visual representation learning
    • Barlow Twins: Self-supervised learning with redundancy reduction
    • VICReg: Variance-Invariance-Covariance Regularization for unsupervised learning
  3. Integration of SSL models with traditional Deep Learning models to improve generalisation in Remote Sensing tasks (30 min)
  4. Challenges, best practices, and evaluation strategies for SSL in Remote Sensing (30 min)
  5. Summary and final thoughts (10 min)
  6. Follow-up tasks and quiz (45–60 min)
Case study: Application of Self-Supervised Learning for crop classification using a limited Remote Sensing dataset from European LUCAS, leveraging SimCLR and MoCo for improved model accuracy

Outcomes:

  • Ability to implement and apply advanced Self-Supervised Learning models such as SimSiam, SimCLR, MoCo, Barlow Twins and VICReg to the European Land Use-Land Cover Area Frame Survey (LUCAS).
  • Understanding of how SSL can be integrated with traditional deep learning models to solve small data challenges
  • Ability to evaluate the effectiveness of SSL approaches in remote sensing tasks such as classification and segmentation.

© 2026 ZALF. All rights reserved.