Related Experiment Video
Updated: Jan 9, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Unlocking Pseudolabel Potential and Alignment for Unpaired Cross-Modality Adaptation in Remote Sensing Image
IEEE Transactions on Neural Networks and Learning Systems
|December 2, 2025
Summary
This study introduces a new method for remote sensing image segmentation, improving how data from different sensors like optical and radar images are used together. The approach enhances model performance when training data is limited across modalities.
Area of Science:
- Remote Sensing
- Computer Vision
- Machine Learning
Background:
- Multisource sensor technology drives the need for multimodal learning in remote sensing (RS) image segmentation.
- Current methods struggle with large-scale paired samples, especially when optical images are annotated but synthetic aperture radar (SAR) images are not.
- This data gap hinders learning discriminative features for SAR images from optical counterparts, necessitating unsupervised domain adaptation (UDA).
Purpose of the Study:
- To propose a novel one-stage adaptation framework, Unlocking Pseudolabel Potential and Alignment (ULPA), for unpaired cross-modality UDA in RS image segmentation.
- To enhance cross-modality knowledge transfer by effectively utilizing available annotations from one modality to adapt models for another.
- To address the challenge of learning from limited or unpaired data across different sensor modalities in RS.
Main Methods:
- Developed the ULPA framework for unpaired cross-modality adaptation in RS image segmentation.
- Employed a Prototypical Multidomain Alignment (PMDA) strategy using contrastive learning between features and prototypes of identical classes across modalities.
- Introduced Unreliable-Sample-Guided Feature Contrast (UFC) loss to improve the utilization of unreliable pixels by separating them based on prediction confidence and using them as negative samples.
Main Results:
- The proposed ULPA framework, integrating PMDA and UFC loss, demonstrated effective cross-modality domain alignment.
- The method significantly boosted the generalization capability of RS image segmentation models.
- Experiments confirmed the enhanced performance in scenarios with unpaired cross-modality data.
Conclusions:
- The ULPA framework offers a robust solution for unpaired cross-modality UDA in remote sensing image segmentation.
- The combination of PMDA and UFC loss effectively bridges the modality gap and improves model generalization.
- This approach facilitates more efficient knowledge transfer in multimodal RS datasets with limited paired annotations.
