HSIGene: A Foundation Model for Hyperspectral Image Generation.

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral imaging (HSI) is crucial for agriculture and environmental monitoring.
• Limited HSI data due to high acquisition costs hinders downstream task performance.
• Existing HSI synthesis methods struggle with reliability, diversity, and controllability.

Purpose of the Study:

• To develop a novel HSI generation foundation model, HSIGene, addressing limitations of current methods.
• To enable precise and reliable HSI generation with multi-condition control.
• To improve spatial diversity and spectral fidelity in synthesized HSIs.

Main Methods:

• Proposed HSIGene, a latent diffusion model with multi-condition control for HSI generation.
• Introduced a data augmentation method using spatial super-resolution to increase training data diversity.
• Developed a two-stage HSI super-resolution framework, including a Rectangular Guided Attention Network (RGAN).

Main Results:

• HSIGene demonstrated the capability to generate a large volume of realistic HSIs.
• The proposed data augmentation and super-resolution methods enhanced spatial diversity while preserving spectral fidelity.
• Generated HSIs proved effective for downstream tasks like denoising and super-resolution.

Conclusions:

• HSIGene offers a significant advancement in controllable and reliable HSI synthesis.
• The integrated data augmentation and super-resolution techniques effectively address HSI data scarcity.
• The model provides a valuable resource for advancing HSI applications in various scientific fields.