Adaptive convolution kernel network for change detection in hyperspectral images.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Feature extraction is crucial for hyperspectral image change detection.
• Varying target sizes and class imbalance pose significant challenges.
• Existing methods struggle with multi-scale features and imbalanced datasets.

Purpose of the Study:

• To develop an improved hyperspectral image change detection method.
• To address challenges of multi-scale target detection and class imbalance.
• To enhance the accuracy and adaptability of change detection algorithms.

Main Methods:

• Proposed an adaptive convolution kernel structure within a U-Net model.
• The adaptive kernel utilizes multiple kernel sizes with learned weights.
• Implemented a weighted cross-entropy loss function to mitigate class imbalance.

Main Results:

• The adaptive convolution kernel effectively extracts multi-scale spatial features.
• The weighted loss function improved detection accuracy for changed pixels.
• Evaluations on four datasets demonstrated superior performance compared to existing methods.

Conclusions:

• The proposed method offers a robust solution for hyperspectral image change detection.
• Adaptive kernels and weighted loss functions are effective for handling scale variation and class imbalance.
• This approach enhances the reliability of change detection in satellite remote sensing imagery.