View in EDS

A multi-task learning framework for dual-polarization SAR imagery despeckling in temporal change detection scenarios.

Saved in:
Bibliographic Details
Title: A multi-task learning framework for dual-polarization SAR imagery despeckling in temporal change detection scenarios.
Authors: Li, Jie1,2,3 (AUTHOR), Shi, Shaowei1 (AUTHOR), Lin, Liupeng1,4 (AUTHOR) linliupeng@whu.edu.cn, Yuan, Qiangqiang1,2 (AUTHOR), Shen, Huanfeng4 (AUTHOR), Zhang, Liangpei5 (AUTHOR)
Source: ISPRS Journal of Photogrammetry & Remote Sensing. Mar2025, Vol. 221, p155-178. 24p.
Subjects: Speckle interference, Convolutional neural networks, Synthetic aperture radar, Deep learning, Transformer models
Abstract: The despeckling task for synthetic aperture radar (SAR) has long faced the challenge of obtaining clean images. Although unsupervised deep learning despeckling methods alleviate this issue, they often struggle to balance despeckling effectiveness and the preservation of spatial details. Furthermore, some unsupervised despeckling approaches overlook the effect of land cover changes when dual-temporal SAR images are used as training data. To address this issue, we propose a multitask learning framework for dual-polarization SAR imagery despeckling and change detection (MTDN). This framework integrates polarization decomposition mechanisms with dual-polarization SAR images, and utilizes a change detection network to guide and constrain the despeckling network for optimized performance. Specifically, the despeckling branch of this framework incorporates polarization and spatiotemporal information from dual-temporal dual-polarization SAR images to construct a despeckling network. It employs various attention mechanisms to recalibrate features across local/global, channel, and spatial dimensions, and before and after despeckling. The change detection branch, which combines Transformer and convolutional neural networks, helps the despeckling branch effectively filter out spatiotemporal information with substantial changes. The multitask joint loss function is weighted by the generated change detection mask to achieve collaborative optimization. Despeckling and change detection experiments are conducted using a dual-polarization SAR dataset to assess the effectiveness of the proposed framework. The despeckling experiments indicate that MTDN efficiently eliminates speckle noise while preserving polarization information and spatial details, and surpasses current leading SAR despeckling methods. The equivalent number of looks (ENL) for MTDN in the agricultural change area increased to 155.0630, and the edge detail preservation (EPD) metric improved to 0.9963, which is better than the comparison methods. Furthermore, the change detection experiments confirm that MTDN yields precise predictions, highlighting its exceptional capability in practical applications. The code, dataset, and pre-trained MTDN will be available at https://github.com/WHU-SGG-RS-Pro-Group/PolSAR-DESPECKLING-MTDN for verification. [ABSTRACT FROM AUTHOR]
Copyright of ISPRS Journal of Photogrammetry & Remote Sensing is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:The despeckling task for synthetic aperture radar (SAR) has long faced the challenge of obtaining clean images. Although unsupervised deep learning despeckling methods alleviate this issue, they often struggle to balance despeckling effectiveness and the preservation of spatial details. Furthermore, some unsupervised despeckling approaches overlook the effect of land cover changes when dual-temporal SAR images are used as training data. To address this issue, we propose a multitask learning framework for dual-polarization SAR imagery despeckling and change detection (MTDN). This framework integrates polarization decomposition mechanisms with dual-polarization SAR images, and utilizes a change detection network to guide and constrain the despeckling network for optimized performance. Specifically, the despeckling branch of this framework incorporates polarization and spatiotemporal information from dual-temporal dual-polarization SAR images to construct a despeckling network. It employs various attention mechanisms to recalibrate features across local/global, channel, and spatial dimensions, and before and after despeckling. The change detection branch, which combines Transformer and convolutional neural networks, helps the despeckling branch effectively filter out spatiotemporal information with substantial changes. The multitask joint loss function is weighted by the generated change detection mask to achieve collaborative optimization. Despeckling and change detection experiments are conducted using a dual-polarization SAR dataset to assess the effectiveness of the proposed framework. The despeckling experiments indicate that MTDN efficiently eliminates speckle noise while preserving polarization information and spatial details, and surpasses current leading SAR despeckling methods. The equivalent number of looks (ENL) for MTDN in the agricultural change area increased to 155.0630, and the edge detail preservation (EPD) metric improved to 0.9963, which is better than the comparison methods. Furthermore, the change detection experiments confirm that MTDN yields precise predictions, highlighting its exceptional capability in practical applications. The code, dataset, and pre-trained MTDN will be available at https://github.com/WHU-SGG-RS-Pro-Group/PolSAR-DESPECKLING-MTDN for verification. [ABSTRACT FROM AUTHOR]
ISSN:09242716
DOI:10.1016/j.isprsjprs.2025.01.030