Section 6

Wrap-up and comparative reflections

Estimated Time: ~5-10 minutes

This practical module brought to life the concepts introduced in your two theoretical lectures by providing guided, hands-on experience with deep learning techniques specifically tailored for agricultural remote sensing under small-data constraints.

You began with a supervised learning approach, training a model on labeled imagery from the LUCAS dataset. You then progressed through three self-supervised learning (SSL) frameworks as MoCo, SimCLR, and SimSiam, where each offering unique ways to extract meaningful representations from unlabeled imagery.

By completing these exercises, you learned how to:

Set up and train a supervised model using PyTorch and labeled Earth Observation data
Apply self-supervised learning techniques that reduce dependence on labeled data
Compare learned representations using tools such as t-SNE, confusion matrices, and accuracy metrics
Reflect on model generalisation, particularly in small and regionally limited datasets

Comparative insights: Supervised vs Self-Supervised Learning

Criteria	Supervised learning	Self-Supervised Learning
Label Requirement	High (must be labeled)	None (pretext task only)
Performance in Small Datasets	Often limited	Generally more robust with limited labels
Transferability	Often task-specific	Higher generalisation to new regions or crops
Training Complexity	Lower (single phase)	Higher (pretraining + optional fine-tuning)
Best Use Case	High-quality labeled datasets	Unlabeled or weakly labeled agricultural imagery

Choosing the right approach.

Scenario	Recommended approach
You have a large, well-annotated dataset	Supervised Learning
You have mostly unlabeled satellite imagery	Self-Supervised Learning (SimCLR, MoCo, SimSiam)
You want to generalize across different regions or seasons	SSL + fine-tuning on a small labeled subset
You are working with few-shot labels or class imbalance	SSL or hybrid approaches with augmentations

Final Thought: In small-data environments, like many real-world agricultural monitoring scenarios (LUCAS dataset), the choice of model architecture is important, but the training strategy is often even more critical. This module has shown that SSL offers a powerful, flexible solution when labeled data is scarce or unevenly distributed. By learning from the data itself, SSL allows for more scalable, transferable, and equitable AI in agriculture.