Multi-spatial resolution hyperspectral image change detection network integrating feature difference structure

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Existing Siamese network-based change detection (CD) methods often rely on Euclidean distance, neglecting deeper CNN-extracted feature disparities.
• Spatial information extraction in CD can be limited by patch-based approaches, potentially misclassifying adjacent pixels or objects.

Purpose of the Study:

• To propose FDS-BINet, a novel multi-spatial resolution network for enhanced change detection in bitemporal hyperspectral images (HSIs).
• To integrate pixel-level and subpixel-level change information for more accurate detection.

Main Methods:

• Developed a Feature Difference Structure (FDS) module for progressively extracting deep, pixel-level differences using residual-enhanced Siamese branches.
• Utilized Bicubic Interpolation (BI) to interpolate the absolute distance spectrum, recovering subpixel-level changes.
• Fused pixel-level and subpixel-level change information within a unified framework.

Main Results:

• The FDS module effectively captures deep difference features from pixel-level spatial information in bitemporal HSIs.
• Bicubic interpolation (BI) enhanced the spatial resolution of absolute distance (AD) spectrum data, enabling subpixel-level feature difference extraction.
• Experimental results show FDS-BINet outperforms representative methods in hyperspectral image change detection.

Conclusions:

• FDS-BINet successfully fuses pixel-level and subpixel-level change information for improved hyperspectral image change detection.
• The proposed method addresses limitations of existing approaches by capturing deeper feature disparities and enabling subpixel analysis.
• FDS-BINet demonstrates superior performance in detecting changes in the Earth's surface from hyperspectral data.