Semantic consistency-aware pseudo-temporal framework for multimodal remote sensing image segmentation.

Area of Science:

• Geospatial analysis
• Computer vision
• Remote sensing

Background:

• Multimodal remote sensing semantic segmentation is crucial for land-cover recognition but struggles with ultra-high-resolution imagery.
• Existing methods suffer from spatial context fragmentation and inconsistent multimodal fusion, limiting generalization.

Purpose of the Study:

• To propose a semantic consistency-aware pseudo-temporal multimodal segmentation framework for ultra-high-resolution remote sensing imagery.
• To address challenges in capturing long-range dependencies and ensuring cross-modal semantic consistency.

Main Methods:

• Developed a framework utilizing the Segment Anything Model (SAM).
• Introduced a pseudo-temporal input construction strategy using random walks and image transformations.
• Implemented a cross-modal temporal interaction module with pyramid fusion and cross-frame attention.
• Employed a prompt-guided decoder with semantic similarity constraints.

Main Results:

• The proposed framework effectively models cross-region contextual dependencies and cross-modal semantic complementarity.
• Achieved significant performance improvements over existing methods on ISPRS Vaihingen and Potsdam datasets.
• Demonstrated enhanced class separability and structural consistency in segmentation results.

Conclusions:

• The semantic consistency-aware pseudo-temporal framework offers a robust solution for multimodal remote sensing segmentation.
• The method shows strong generalization capabilities in complex environments and ultra-high-resolution imagery.
• Validated effectiveness for fine-grained land-cover recognition and geospatial analysis.