Advancing perturbation space expansion based on information fusion for semi-supervised remote sensing image semantic segmentation

• Published in: Information fusion 2025-05, Vol.117, p.102830, Article 102830
• Main Authors: Zhou, Liang, Duan, Keyi, Dai, Jinkun, Ye, Yuanxin
• Format: Article
• Language: English
• Subjects: Information fusion; Perturbation space expansion; Remote sensing image segmentation; Semi-supervised semantic segmentation
• ISSN: 1566-2535
• DOI: 10.1016/j.inffus.2024.102830

Existing deep models have greatly enhanced the performance of semantic segmentation in remote sensing (RS) images, but they are often limited by the scarcity of labeled samples. Semi-supervised learning (SSL) can leverage a vast amount of unlabeled data for model training, effectively overcoming the reliance on labeled data. Nonetheless, due to the complexity and diversity of ground objects in RS images, existing perturbation methods designed for natural images exhibit significant limitations when extended to remote sensing imagery, failing to construct a broader perturbation space. Incorporating the unique characteristics of RS images, this work presents a novel consistency regularization semi-supervised framework, named FusionMatch, which innovatively integrates multi-modal near-infrared information (termed as NIRPerb) and applies differentiated pan-sharpening fusion techniques (termed as PSPerb) during the perturbation process, thereby significantly expanding the perturbation space. FusionMatch not only enriches the dataset with additional spectral dimensions but also improves the model’s capacity to discern subtle variations and patterns. Moreover, FusionMatch can be seamlessly integrated with other consistency regularization methods to enhance the capabilities of segmentation models. Two binary classification datasets and two multi-class datasets are used to assess the segmentation performance of FusionMatch. The experimental results confirm the robustness of FusionMatch compared to other leading semi-supervised approaches. [Display omitted] •Propose a semi-supervised segmentation method with information fusion perturbation.•Expand the perturbation space by incorporating near-infrared band information.•Achieve a broader perturbation space by using differentiated pan-sharpening methods.•Our method can be seamlessly integrated with other semi-supervised frameworks.