Unsupervised Domain Adaptation for Land Cover Mapping
Overview
This research enhances land cover mapping using satellite imagery through unsupervised domain adaptation, addressing the challenge of limited labeled data in target regions.
Original research can be found here
Problem Statement
- Land cover mapping is crucial but labeling data for every region is resource-intensive
- Need to transfer knowledge from labeled (source) to unlabeled (target) datasets effectively
Approach
1. Dataset
- Utilized portion of “Five-Billion-Pixels” dataset
- High-resolution satellite images with 24 land cover categories
2. Model Architecture
- U-Net based architecture for semantic segmentation
3. Unsupervised Domain Adaptation
- Siamese network with two branches (source and target domains)
- Adapts model to unlabeled target data
4. Dynamic Pseudo-Labeling
- Gradually increases pseudo-labeled target pixels over training epochs
5. Key Improvements
- Advanced loss functions (Dice Loss, Combined Focal-Dice Loss)
- Extended data augmentation techniques
- Patching strategy for high-resolution images
Results
- Tested by training on one city (source) and predicting on different cities (target)
- Promising results on small subsets, especially with new loss functions and data augmentation
Goal
Develop an effective, adaptable method for land cover mapping across different geographical regions with limited labeled data.
The project’s code and detailed results are available on GitHub.