Transfer Learning
A machine learning technique where a model trained on one task is adapted for a different but related task. In geospatial ML, models pre-trained on large satellite image datasets can be fine-tuned for specific applications with limited labeled data, dramatically reducing training time and data requirements.
Overview
Transfer learning adapts a model pre-trained on a large dataset to a new, smaller target dataset. In earth observation, this addresses the scarcity of labeled satellite imagery — using visual features from millions of images and fine-tuning on hundreds or thousands of labeled samples.
How It Works
A backbone network pre-trained on a source dataset has its task-specific head replaced. First, backbone weights are frozen and only the new head is trained. Then the entire model is fine-tuned end-to-end with a lower learning rate. Source datasets can be general (ImageNet) or domain-specific (SatlasPretrain with 302 million labels from Sentinel-2/NAIP).
Key Facts
- Domain-specific pre-training on satellite data consistently outperforms ImageNet for EO tasks.
- Can reduce need for labeled data by 10x or more compared to training from scratch.
- SatlasPretrain provides 302 million labels across 137 categories for EO pre-training.
- The domain gap between natural photographs and satellite imagery is a key challenge.
Applications
Few-Shot Land Cover Mapping
Fine-tuning pre-trained models on small labeled datasets from new regions.
Damage Assessment
Adapting models to rapidly classify damage levels from post-disaster imagery.
Cross-Sensor Adaptation
Transferring models between different satellite sensors with minimal retraining.
Limitations & Considerations
ImageNet features transfer imperfectly to EO due to viewpoint differences. Multispectral and SAR data have no natural-image equivalent. Models may encode geographic biases. Fine-tuning on very small datasets risks overfitting.
History & Background
Became practical with deep CNNs on ImageNet (2012). Early EO applications in 2015-2016. Domain-specific datasets emerged around 2020. Has evolved into the foundation model paradigm.
Related Terms
Analyze Transfer Learning data with LYRASENSE
Use our agentic notebook environment to work with satellite data and compute indices like Transfer Learning — no setup required.