LaVIDE: Language-Prompted Satellite Change Detection via Map-Image Alignment.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Change detection in remote sensing is crucial for monitoring Earth's surface.
• A significant challenge is the semantic gap between map data and image details, hindering accurate change detection.
• Existing methods struggle with pixel-level comparisons or segmentation errors.

Purpose of the Study:

• To propose a novel framework, LaVIDE (Language-VIsion Discriminator for dEtecting changes), to bridge the semantic gap in map-image change detection.
• To enhance the robustness and accuracy of change detection by leveraging language as an intermediary.
• To facilitate rapid map updating with minimal human intervention.

Main Methods:

• Developed a framework using language as an intermediary to connect high-level map semantics with low-level image details.
• Introduced restricted prompt learning for context-aware textual prompts aligning map semantics and image content.
• Implemented an object-aware embedding enhancement strategy to integrate object-level attributes into map representations.

Main Results:

• LaVIDE achieved significant performance improvements over state-of-the-art methods on four benchmarks (DynamicEarthNet, HRSCD, BANDON, SECOND).
• Demonstrated an 18.4% improvement in IoU for multi-class change detection and a 5.2% improvement for single-class change detection.
• Validated the framework's effectiveness in advancing map-image change detection accuracy.

Conclusions:

• The LaVIDE framework successfully bridges the semantic gap in remote sensing change detection using a language-vision approach.
• The method offers a practical solution for efficient map updating with broad applications in urban planning, disaster assessment, and ecological conservation.
• The study advances the field by enabling more accurate and automated Earth surface monitoring.