Prototype-Based Pseudo-Label Denoising for Source-Free Domain Adaptation in Remote Sensing Semantic Segmentation

2025-09-21

AI Summary

In source-free domain adaptation (SFDA) for remote sensing image semantic segmentation, the absence of ground-truth labels in the target domain leads to severe pseudo-label noise and hinders effective domain shift (DS) mitigation. To address this, we propose ProSFDA, a prototype-guided self-training framework. Its core contributions are: (i) dynamic pseudo-label weighting based on class prototypes to suppress noise propagation; and (ii) prototype-level contrastive learning to enforce intra-class feature compactness and inter-class separation, thereby enhancing discriminative representation learning in the target domain. ProSFDA operates solely with a pre-trained source model and unlabeled target data—requiring no access to source data or additional target annotations. Extensive experiments on benchmark remote sensing datasets (e.g., LoveDA, Vaihingen) demonstrate that ProSFDA consistently outperforms state-of-the-art SFDA methods, achieving absolute mIoU gains of 3.2–5.7 percentage points, effectively alleviating domain shift and improving segmentation robustness.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Segmentation
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Security and Privacy: Data transparency and provenance

Abstract

Source-Free Domain Adaptation (SFDA) enables domain adaptation for semantic segmentation of Remote Sensing Images (RSIs) using only a well-trained source model and unlabeled target domain data. However, the lack of ground-truth labels in the target domain often leads to the generation of noisy pseudo-labels. Such noise impedes the effective mitigation of domain shift (DS). To address this challenge, we propose ProSFDA, a prototype-guided SFDA framework. It employs prototype-weighted pseudo-labels to facilitate reliable self-training (ST) under pseudo-labels noise. We, in addition, introduce a prototype-contrast strategy that encourages the aggregation of features belonging to the same class, enabling the model to learn discriminative target domain representations without relying on ground-truth supervision. Extensive experiments show that our approach substantially outperforms existing methods.

Problem

• Addresses noisy pseudo-labels in source-free domain adaptation
• Mitigates domain shift in remote sensing image segmentation
• Enables discriminative target domain learning without ground-truth

Innovation

• Prototype-weighted pseudo-labels for reliable self-training
• Prototype-contrast strategy to aggregate same-class features
• Learning discriminative target representations without ground-truth supervision