A Multistage Information Complementary Fusion Network Based on Flexible-Mixup for HSI-X Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Mixup data augmentation aids model regularization, particularly in data-scarce remote sensing.
• Existing Mixup methods can negatively impact label space due to inconsistent augmentation ratios.
• Multimodal interaction strategies struggle with diverse remote sensing data combinations.

Purpose of the Study:

• To propose a novel method for hyperspectral-X image classification.
• To address limitations of standard Mixup augmentation in remote sensing.
• To effectively fuse diverse remote sensing data modalities.

Main Methods:

• Introduced a flexible-mixup (FlexMix) strategy to align mixed images with labels by adjusting label weights.
• Developed a multistage information complementary fusion network (MCFNet) for hyperspectral and X-modal data.
• Extracted and fused features from hyperspectral, multispectral, SAR, and LiDAR data through multiple interaction stages.

Main Results:

• Flex-MCFNet effectively expands training data for hyperspectral-X image classification.
• The proposed method adequately regularizes models across different data combinations.
• Achieved state-of-the-art performance in hyperspectral-X image classification tasks.

Conclusions:

• The FlexMix strategy mitigates negative impacts on the label space during augmentation.
• MCFNet successfully integrates and enhances information from diverse remote sensing modalities.
• Flex-MCFNet offers a robust solution for remote sensing image classification with limited and multimodal data.