Synthetic AI-Generated Satellite Imagery to Improve Earth Observation-Based Neural Networks

Area of Science:

• Earth Observation
• Computer Vision
• Artificial Intelligence

Background:

• Acquiring labeled satellite imagery for Earth observation is expensive and time-consuming.
• Real satellite images often contain defects like cloud cover and misaligned channels.
• Generative AI can create realistic synthetic data to address data scarcity and defects.

Purpose of the Study:

• To demonstrate the feasibility of transferring knowledge from AI-generated datasets to Earth observation missions.
• To evaluate the effectiveness of synthetic data augmentation for satellite image segmentation tasks.

Main Methods:

• Utilized Pix2Pix, CUT, and ControlNet models to synthesize satellite imagery from Spanish map tiles.
• Trained identical U-Net instances on both real and synthetic datasets for building, road, and water segmentation.
• Tested the trained models on independent authentic satellite imagery to assess generalizability.

Main Results:

• Synthetic data incorporation improved segmentation performance compared to using real data alone.
• Maximum Dice scores increased by 0.9% for buildings (to 54.1%), 2.3% for roads (to 38.6%), and 4.1% for waterbodies (to 46.5%).
• A decoupling between visual realism and functional utility of synthetic data was observed.

Conclusions:

• Structural-guided synthetic data augmentation is a viable strategy for enhancing Earth observation tasks.
• This approach offers a robust and adaptable solution for diverse sensors and segmentation objectives.
• AI-generated data can effectively supplement real-world data, overcoming limitations in quantity and quality.