PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features

Geometric distortions and significant nonlinear radiometric differences in multimodal remote sensing images (MRSIs) introduce substantial noise in feature extraction. Single-branch convolutional neural networks fail to capture global image features and integrate local and global information effectiv...

Full description

Saved in:
Bibliographic Details
Main Authors: Haiqing He, Shixun Yu, Yongjun Zhang, Yufeng Zhu, Ting Chen, Fuyang Zhou
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10999084/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850232654641758208
author Haiqing He
Shixun Yu
Yongjun Zhang
Yufeng Zhu
Ting Chen
Fuyang Zhou
author_facet Haiqing He
Shixun Yu
Yongjun Zhang
Yufeng Zhu
Ting Chen
Fuyang Zhou
author_sort Haiqing He
collection DOAJ
description Geometric distortions and significant nonlinear radiometric differences in multimodal remote sensing images (MRSIs) introduce substantial noise in feature extraction. Single-branch convolutional neural networks fail to capture global image features and integrate local and global information effectively, yielding deep descriptors with low discriminability and limited robustness. Moreover, the lack of comprehensive training data further limits the network's performance, which poses a formidable challenge to existing matching methods in securing adequate and evenly distributed corresponding points. This article proposes a novel method called phase-integrated aggregated deep feature matching (PI-ADFM), designed to address these challenges. Initially, a phase structure feature detector is introduced, which amalgamates the structural attributes and phase information of images to distill keypoints that are highly repeatable and exhibit minimal redundancy. Subsequently, an attention-based multilevel feature interaction and aggregation module is crafted to encapsulate a comprehensive representation of both local and global features of keypoints. This is followed by the integration of a dense feature fusion module designed to sift through and amalgamate key features, thereby capturing highly discriminative deep semantic features that serve as descriptors for similarity measures. Finally, a multilevel outlier removal strategy is proposed to effectively reduce mismatches. Experimental results substantiate that, in juxtaposition with state-of-the-art methods, the PI-ADFM method has significantly augmented the count of matches for optical-infrared and optical synthetic aperture radar images by a factor of at least 1.7 and 3.7, respectively, while concurrently enhancing the accuracy by a minimum of 10% and 6%, respectively. These enhancements markedly bolster the robustness and reliability of MRSI matching endeavors.
format Article
id doaj-art-80fa3a545896483c81f2f0876b914d9c
institution OA Journals
issn 1939-1404
2151-1535
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
spelling doaj-art-80fa3a545896483c81f2f0876b914d9c2025-08-20T02:03:07ZengIEEEIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing1939-14042151-15352025-01-0118131921321110.1109/JSTARS.2025.356917410999084PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep FeaturesHaiqing He0https://orcid.org/0000-0001-9361-0219Shixun Yu1https://orcid.org/0009-0001-5835-0949Yongjun Zhang2https://orcid.org/0000-0001-9845-4251Yufeng Zhu3Ting Chen4Fuyang Zhou5https://orcid.org/0009-0005-5394-0762National Key Laboratory of Uranium Resources Exploration-Mining and Nuclear Remote Sensing, East China University of Technology, Nanchang, ChinaNational Key Laboratory of Uranium Resources Exploration-Mining and Nuclear Remote Sensing, East China University of Technology, Nanchang, ChinaSchool of Remote Sensing and Information Engineering, Wuhan University, Wuhan, ChinaNational Key Laboratory of Uranium Resources Exploration-Mining and Nuclear Remote Sensing, East China University of Technology, Nanchang, ChinaSchool of Water Resources and Environmental Engineering, East China University of Technology, Nanchang, ChinaNational Key Laboratory of Uranium Resources Exploration-Mining and Nuclear Remote Sensing, East China University of Technology, Nanchang, ChinaGeometric distortions and significant nonlinear radiometric differences in multimodal remote sensing images (MRSIs) introduce substantial noise in feature extraction. Single-branch convolutional neural networks fail to capture global image features and integrate local and global information effectively, yielding deep descriptors with low discriminability and limited robustness. Moreover, the lack of comprehensive training data further limits the network's performance, which poses a formidable challenge to existing matching methods in securing adequate and evenly distributed corresponding points. This article proposes a novel method called phase-integrated aggregated deep feature matching (PI-ADFM), designed to address these challenges. Initially, a phase structure feature detector is introduced, which amalgamates the structural attributes and phase information of images to distill keypoints that are highly repeatable and exhibit minimal redundancy. Subsequently, an attention-based multilevel feature interaction and aggregation module is crafted to encapsulate a comprehensive representation of both local and global features of keypoints. This is followed by the integration of a dense feature fusion module designed to sift through and amalgamate key features, thereby capturing highly discriminative deep semantic features that serve as descriptors for similarity measures. Finally, a multilevel outlier removal strategy is proposed to effectively reduce mismatches. Experimental results substantiate that, in juxtaposition with state-of-the-art methods, the PI-ADFM method has significantly augmented the count of matches for optical-infrared and optical synthetic aperture radar images by a factor of at least 1.7 and 3.7, respectively, while concurrently enhancing the accuracy by a minimum of 10% and 6%, respectively. These enhancements markedly bolster the robustness and reliability of MRSI matching endeavors.https://ieeexplore.ieee.org/document/10999084/Deep learningfeature interaction and aggregationfeature matchingmultimodal imagesphase structure
spellingShingle Haiqing He
Shixun Yu
Yongjun Zhang
Yufeng Zhu
Ting Chen
Fuyang Zhou
PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Deep learning
feature interaction and aggregation
feature matching
multimodal images
phase structure
title PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features
title_full PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features
title_fullStr PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features
title_full_unstemmed PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features
title_short PI-ADFM: Enhancing Multimodal Remote Sensing Image Matching Through Phase-Integrated Aggregated Deep Features
title_sort pi adfm enhancing multimodal remote sensing image matching through phase integrated aggregated deep features
topic Deep learning
feature interaction and aggregation
feature matching
multimodal images
phase structure
url https://ieeexplore.ieee.org/document/10999084/
work_keys_str_mv AT haiqinghe piadfmenhancingmultimodalremotesensingimagematchingthroughphaseintegratedaggregateddeepfeatures
AT shixunyu piadfmenhancingmultimodalremotesensingimagematchingthroughphaseintegratedaggregateddeepfeatures
AT yongjunzhang piadfmenhancingmultimodalremotesensingimagematchingthroughphaseintegratedaggregateddeepfeatures
AT yufengzhu piadfmenhancingmultimodalremotesensingimagematchingthroughphaseintegratedaggregateddeepfeatures
AT tingchen piadfmenhancingmultimodalremotesensingimagematchingthroughphaseintegratedaggregateddeepfeatures
AT fuyangzhou piadfmenhancingmultimodalremotesensingimagematchingthroughphaseintegratedaggregateddeepfeatures