CrossEarth: Geospatial Vision Foundation Model for Domain Generalizable Remote Sensing Semantic Segmentation.

Area of Science:

• Earth and Space Sciences
• Computer Science
• Artificial Intelligence

Background:

• Remote Sensing (RS) images exhibit significant domain gaps due to variations in location, wavelength, and sensor type.
• Existing cross-domain methods often focus on Domain Adaptation (DA) for predefined domains, not generalization to unseen scenarios.
• Research in Remote Sensing Domain Generalization (RSDG) for semantic segmentation is limited, with current models underperforming on diverse, unseen domains.

Purpose of the Study:

• To introduce the first vision foundation model specifically designed for Remote Sensing Domain Generalization (RSDG) semantic segmentation.
• To address the limitations of current models in generalizing to diverse and unseen remote sensing scenarios.
• To establish a comprehensive benchmark for evaluating RSDG semantic segmentation models.

Main Methods:

• Developed CrossEarth, a novel vision foundation model for RSDG semantic segmentation.
• Implemented a data-level Earth-Style Injection pipeline to enhance domain variability.
• Utilized a model-level Multi-Task Training pipeline to improve generalization capabilities.
• Curated a new RSDG benchmark with 32 semantic segmentation scenarios across diverse conditions.

Main Results:

• CrossEarth demonstrated strong cross-domain generalization capabilities.
• The proposed Earth-Style Injection and Multi-Task Training pipelines significantly improved performance.
• Extensive experiments on the new benchmark confirmed CrossEarth's superiority over state-of-the-art methods.
• The benchmark provides a robust evaluation for future RSDG model development.

Conclusions:

• CrossEarth is the first effective vision foundation model for RSDG semantic segmentation.
• The model's architecture and training strategies enable superior generalization across diverse remote sensing domains.
• The newly curated benchmark facilitates comprehensive evaluation and advancement of RSDG research.