Lesson 4

Deep Learning for Agricultural Remote Sensing under Small Data Constraints (DL4Agro)

Objectives:

Build foundational and practical knowledge of modern DL approaches for small-data challenges in remote sensing. Emphasis is placed on supervised and self-supervised learning methods relevant to agricultural monitoring and land use applications.

Content:

• Theoretical background on DL models for small and weakly labeled data
• Overview of key strategies: transfer learning, self-supervised learning (SimCLR, SimSiam, MoCo), spatial validation
• Real-world applications using LUCAS dataset for land cover and crop classification
• Comparative analysis of supervised vs. self-supervised methods
• Insights on model generalisation, feature representation, and practical deployment

Activities:

• Recorded lectures and concept walkthroughs
• Practical coding sessions using Jupyter Notebooks (supervised + 3 SSL models)
• Visualizations and evaluation of learned features
• Reflections on model behavior in low-data environments

Assessments:

• Multiple-choice quizzes on lecture content
• Matching tasks and short open-answer questions
• Optional coding assignments and reflection prompts (optional)

Resources:

• Prepared image datasets from the LUCAS survey
• Sample notebooks for supervised and SSL methods
• Code templates for training and testing models
• Guides for model visualization and performance evaluation
• Access to Python packages: torch, torchvision, lightly, matplotlib, scikit-learn

Outcomes:

• Understand the small data problem in remote sensing
• Apply both supervised and self-supervised models to labeled/unlabeled satellite data
• Adapt deep learning workflows to agricultural datasets
• Evaluate and interpret model performance using visual and statistical tools
• Gain practical experience transferable to real agricultural monitoring systems

WP4: Deep Learning Approaches to Address Small-Data Problems in Agricultural Remote Sensing

Authors:

Dr. Anastasiia Safonova (Contact: )
Prof. Dr. Masahiro Ryo