Spatial Channel Attention based Change Detection in Synthetic Aperture Radar Images

Change Detection (CD) in Synthetic Aperture Radar (SAR) images is a key research field in remote sensing. Recently, CD in SAR images has received more attention due to the availability of SAR images, irrespective of weather conditions. However, it is difficult to exploit the changed information from...

Full description

Saved in:
Bibliographic Details
Published in2023 International Conference on Device Intelligence, Computing and Communication Technologies, (DICCT) pp. 537 - 541
Main Authors Khandelwal, Luv, Sujit Vudattu, V N, Chandra Pati, Umesh
Format Conference Proceeding
LanguageEnglish
Published IEEE 17.03.2023
Subjects
Change detection
convolution neural network
Neural networks
Performance evaluation
Radar detection
Radar imaging
Radar polarimetry
spatial channel attention
Speckle
synthetic aperture radar
Training
Online AccessGet full text

Cover

More Information
Summary:Change Detection (CD) in Synthetic Aperture Radar (SAR) images is a key research field in remote sensing. Recently, CD in SAR images has received more attention due to the availability of SAR images, irrespective of weather conditions. However, it is difficult to exploit the changed information from SAR images, as these are subject to speckle noise. In this work, a Spatial Channel Attention (SCA) based CD network (SCACDNet) has been proposed. The SCACDNet is designed by integrating convolution neural network with the SCA module. Initially, the two SAR images are preprocessed and then, difference image is produced using the Log-Ratio operator. Then, the hierarchical fuzzy c-means clustering is used to classify the pixels into unchanged, changed and fuzzy classes. Now, image patches around highly probable unchanged and changed pixels are fed as an input to the network. After training, the pixels belong to fuzzy class are classified by the SCACDNet to produce a change map. The quantitative and qualitative results obtained for three SAR datasets demonstrate the superior performance of the proposed model.
DOI:10.1109/DICCT56244.2023.10110172