Style transfer-based domain adaptation for vegetation segmentation with optical imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Domain adaptation in optical imagery is crucial for deep learning models using diverse sensor systems.
• Cycle-consistent adversarial networks are effective for dataset transformation but risk losing vital spectral information.
• Optical airborne and spaceborne sensors are vital for monitoring vegetation health and coverage.

Purpose of the Study:

• To develop a novel style transfer method that preserves spectral information essential for vegetation analysis.
• To improve the performance of deep learning models in segmenting vegetation areas across different optical datasets.
• To address the loss of characteristic spectral information during image style transfer.

Main Methods:

• A cycle-consistent adversarial domain adaptation method with four input channels was developed.
• The method incorporates index-based metrics for vegetation segmentation.
• The approach focuses on preserving the ratio between near-infrared (NIR) and RGB bands.

Main Results:

• The proposed method successfully preserves the spectral ratio between NIR and RGB bands.
• Significant improvements in vegetation segmentation network performance were observed in the target domain.
• The method demonstrates effective domain adaptation while retaining critical spectral characteristics.

Conclusions:

• The developed cycle-consistent adversarial domain adaptation method enhances vegetation segmentation by preserving essential spectral information.
• This approach offers a robust solution for adapting optical imagery from different sensors for vegetation analysis.
• The findings contribute to more accurate and reliable vegetation monitoring using remote sensing data.