DNN-Pol SAR: Urban Image Segmentation and Classification using Polarimetric SAR based on DNNs
Soumyadip Sarkar¹, Farhan Hai Khan², Shobhit Kumar³, Tamesh Halder⁴, Dipjyoti Paul⁵, Debashish Chakravarty⁶

¹Soumyadip Sarkar, Institute of Engineering & Management, Kolkata, India.

²Farhan Hai Khan, Institute of Engineering & Management, Kolkata, India.

³Shobhit Kumar, Institute of Engineering & Management, Kolkata, India.

⁴Tamesh Halder, Department of Mining Engineering, Indian Institute of Technology, Kharagpur, India.

⁵Dipjyoti Paul, Department of Computer Science, University of Crete, Heraklion, Crete, Greece.

⁶Debashish Chakravarty, Department of Mining Engineering, Indian Institute of Technology, Kharagpur, India. 

Manuscript received on 12 April 2024 | Revised Manuscript received on 11 May 2024 | Manuscript Accepted on 15 May 2024 | Manuscript published on 30 May 2024 | PP: 1-13 | Volume-11 Issue-5, May 2024 | Retrieval Number: 100.1/ijies.E444813050624 | DOI: 10.35940/ijies.E4448.11050524

Open Access | Editorial and Publishing Policies | Cite | Zenodo | OJS | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Synthetic Aperture Radar (SAR) image segmentation and classification are popular techniques for learning and detecting objects such as buildings, trees, monuments, crops, water bodies, hills, etc. The SAR technique is being utilised for urban development and city planning, building control of municipal objects, identifying optimal locations, and detecting changes in existing systems, among other applications, by leveraging polarimetry based on Deep Neural Networks. In this paper, we propose a technique for urban image segmentation and Classification using Polarimetric SAR based on Deep Neural Networks (DNN-PolSAR). In our proposed DNN-PolSAR technique, we utilise Mask-RCNN, LinkNet, FPN, and PSP-Net as model architectures, while ResNet-50, ResNet-101, ResNet-152, and VGG-19 are employed as backbone networks.We first apply polarimetric decomposition to airborne Uninhabited Aerial Vehicle Synthetic Aperture (UAVSAR) images of urban areas, and then the decomposed images are fed to DNNs for segmentation and classification. We then simulate DNN-PolSAR considering different hyperparameters and compare the obtained scores of these hyperparameters against the used model architectures and backbone networks. In comparison, it is found that DNN-PolSAR, based on the FPN model with ResNet152, performed the best for segmentation and classification. The mean Average Precision (mAP) score of the DNN-PolSAR based on FPN with a pixel accuracy of 90.9% is 0.823, which outperforms other Deep Learning models.

Keywords: Polarimetric SAR, FPN, PSPNet, Mask-RCNN, LinkNet, Image Segmentation.
Scope of the Article: Classification