Three-band SWIR imaging for Pix2Pix-based colorization

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Short-wave infrared (SWIR) imaging is crucial for remote sensing but lacks color information, hindering analysis.
• Deep learning methods struggle with spectral ambiguity in SWIR data, making color reconstruction ill-posed.

Purpose of the Study:

• To develop a novel framework for reconstructing visible colors from SWIR imagery.
• To address the challenge of spectral ambiguity and improve human interpretability of SWIR data.

Main Methods:

• A physics-data dual-driven band selection framework integrated with a Pix2Pix network.
• Three stages: physical pre-selection, statistical coarse-ranking (modified optimum index factor), and generative evaluation.
• Identified an optimal three-band combination (1000-1050, 1050-1100, 1525-1575 nm) and used a ResNet generator.

Main Results:

• The proposed method significantly outperforms standard approaches in color reconstruction from SWIR data.
• Achieved an 8.63% improvement in PSNR, 18.87% in SSIM, and 31.28% in LPIPS compared to a single-band U-Net baseline.
• Effectively decoupled mapping ambiguities by filtering radiometric noise.

Conclusions:

• The physics-data dual-driven framework successfully reconstructs color information in SWIR imagery.
• This approach enhances scene analysis and overcomes limitations of existing deep learning methods for SWIR data.
• The identified optimal band combination and generative approach offer a robust solution for SWIR image interpretation.