DI3CL: Contrastive Learning With Dynamic Instances and Contour Consistency for SAR Land-Cover Classification...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Supervised learning methods for SAR land-cover classification require extensive labeled datasets, limiting scalability and adaptability.
• Existing approaches struggle with generalization across diverse application scenarios and geographic regions.
• A need exists for a general-purpose foundation model to streamline SAR land-cover classification development.

Purpose of the Study:

• To develop a general-purpose foundation model for SAR land-cover classification.
• To introduce a novel pre-training framework, Dynamic Instance and Contour Consistency Contrastive Learning (DI3CL), to enhance model performance.
• To improve the robustness and generalization capabilities of SAR land-cover classification models.

Main Methods:

• Developed a Dynamic Instance and Contour Consistency Contrastive Learning (DI3CL) pre-training framework.
• Incorporated a Dynamic Instance (DI) module for enhanced global contextual awareness and a Contour Consistency (CC) module for improved structural discrimination.
• Constructed a large-scale dataset (SARSense) with 460,532 SAR images for comprehensive feature capture.

Main Results:

• The DI3CL framework demonstrated superior performance compared to existing methods in extensive experiments.
• The foundation model showed strong generalization capabilities across various SAR land-cover classification tasks, including mapping, water detection, and road extraction.
• Pre-trained weights and code are publicly available, facilitating further research and application.

Conclusions:

• The proposed DI3CL foundation model effectively addresses the limitations of supervised learning in SAR land-cover classification.
• The DI3CL framework significantly enhances model robustness, generalization, and structural discrimination.
• The developed foundation model serves as a robust cornerstone for accelerating diverse downstream SAR land-cover classification applications.