General Hyperspectral Image Super-Resolution via Meta-Transfer Learning.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning significantly advanced hyperspectral super-resolution (SR).
• Data scarcity, high dimensionality, and domain differences hinder deep model generalizability in hyperspectral image (HSI) SR.
• Existing methods struggle with stable pattern learning and cross-dataset applicability.

Purpose of the Study:

• To develop a general hyperspectral SR framework addressing data limitations and domain differences.
• To enhance the generalizability and performance of deep learning models for HSI SR.
• To provide a versatile framework applicable to existing SR methods.

Main Methods:

• Utilized meta-transfer learning (MTL) by randomly sampling spectral ranges for diverse task experience accumulation.
• Implemented a task schedule to progressively increase spectral bands, bridging domain gaps between HSI datasets.
• Designed an advanced network architecture incorporating multifusion features.

Main Results:

• Achieved superior qualitative and quantitative performance in hyperspectral SR.
• Demonstrated robust adaptability to new and challenging HSI samples with minimal training.
• Showcased improved generalizability across various datasets and circumstances.

Conclusions:

• The proposed MTL framework significantly enhances hyperspectral SR performance and generalizability.
• The method effectively mitigates challenges posed by data scarcity and domain differences in HSI datasets.
• This general framework offers a promising approach for advancing hyperspectral SR research and applications.