Day 3: Deep Learning for Earth Observation
CNNs, U-Net, and Transfer Learning for Satellite Imagery
Day 3: Deep Learning for Earth Observation
CNNs, U-Net, and Transfer Learning for Satellite Imagery
Leveraging neural networks for advanced EO image analysis
Day 3 Overview
Day 3 explores advanced deep learning techniques for Earth Observation, focusing on semantic segmentation and object detection. You’ll master the U-Net architecture for pixel-wise classification tasks like flood mapping, and learn modern object detection methods (YOLO, Faster R-CNN) for identifying and counting objects in satellite imagery. Hands-on sessions use Philippine case studies including Central Luzon flood mapping and Metro Manila urban monitoring.
✅ Session 1: Semantic Segmentation with U-Net - Complete ✅ Session 2: Flood Mapping Lab (U-Net + SAR) - Complete ✅ Session 3: Object Detection Techniques - Complete ✅ Session 4: Object Detection Lab (Transfer Learning) - Complete
All Day 3 sessions are now ready for delivery!
Learning Objectives
By the end of Day 3, you will be able to:
- Understand U-Net encoder-decoder architecture and skip connections for semantic segmentation
- Explain the differences between semantic segmentation, instance segmentation, and object detection
- Implement U-Net models from scratch using TensorFlow/Keras for pixel-wise classification
- Process Sentinel-1 SAR imagery for flood detection and disaster response applications
- Train semantic segmentation models with dice loss and IoU metrics
- Apply U-Net to real-world flood mapping scenarios in Central Luzon
- Comprehend object detection architectures (YOLO, Faster R-CNN, RetinaNet)
- Calculate object detection metrics including mAP, IoU, precision, and recall
- Implement transfer learning with pre-trained backbone networks (ResNet, EfficientNet)
- Fine-tune pre-trained object detection models for building footprint extraction
- Design data augmentation strategies appropriate for EO tasks
- Handle imbalanced datasets using weighted loss functions and sampling strategies
- Evaluate segmentation quality using confusion matrices, IoU, and F1-scores
- Visualize model predictions, feature maps, and attention mechanisms
- Deploy operational workflows for disaster mapping and urban monitoring
Today’s Schedule
| Time | Session | Topic | Materials |
|---|---|---|---|
| 09:00-10:30 | 1 | Semantic Segmentation with U-Net | Theory + Demos |
| 10:30-13:00 | 2 | Flood Mapping Lab (Central Luzon) | Hands-on Lab |
| 14:00-15:30 | 3 | Object Detection Techniques for EO | Theory + Case Studies |
| 15:30-18:00 | 4 | Building Detection Lab (Metro Manila) | Hands-on Lab |
Training Sessions
Session 1
Semantic Segmentation with U-Net
- U-Net architecture fundamentals
- Encoder-decoder networks
- Skip connections explained
- Semantic vs instance segmentation
- EO applications
Session 2
Flood Mapping Lab (Central Luzon)
- Sentinel-1 SAR flood mapping
- U-Net implementation in TensorFlow
- Training pipeline development
- Performance evaluation
- Operational deployment
Session 3
Object Detection Techniques for EO
- R-CNN family overview
- YOLO architecture
- Object detection metrics (mAP, IoU)
- Transfer learning strategies
- Philippine use cases
Session 4
Building Detection Lab (Metro Manila)
- Transfer learning with ResNet backbone
- Fine-tuning pre-trained models
- Building footprint detection
- Model evaluation and visualization
- Deployment considerations
Prerequisites
From Previous Days
Before Day 3, you should have completed:
Technical Setup
What’s Next?
After Day 3, you’ll progress to:
Day 4: Advanced Topics & Capstone Projects - Time series analysis, multi-modal data fusion, foundation models, and hands-on project work.
Quick Links
Session 1: Semantic Segmentation Session 2: Flood Mapping Lab Session 3: Object Detection Session 4: Building Detection Lab Slides: Session 1 Slides: Session 2 Slides: Session 3 Slides: Session 4 Data Acquisition Guide Back to Course Home
Sessions 2 & 4 use synthetic/demo data for immediate execution and hands-on learning. This approach:
- ✅ Allows you to run labs immediately (no downloads or preprocessing)
- ✅ Focuses on deep learning methodology and workflows
- ✅ Teaches production-ready code (same code works with real data)
- ✅ Matches industry best practices (e.g., Kaggle, academic courses)
For production applications: See the Data Acquisition Guide for obtaining real Sentinel-1 SAR and Sentinel-2 optical data, including Google Earth Engine scripts and annotation tools.
Day 3 is part of the CoPhil 4-Day Advanced Training on AI/ML for Earth Observation, funded by the European Union under the Global Gateway initiative and delivered in partnership with PhilSA and DOST.