URS: An Unsupervised Radargram Segmentation Network Based on Self-Supervised ViT With Contrastive Feature Learning Framework

Abstract

Radar sounders are air and space-borne nadir-looking sensors operating in high-frequency (HF) or very high-frequency (VHF) bands and collect subsurface backscattered returns by transmitting electromagnetic pulses. The backscatter echoes are coherently integrated to generate radargrams for investigating and identifying geophysical characteristics of subsurface targets. While recent efforts have been made to develop supervised or semisupervised deep learning models for segmenting radargrams, obtaining accurate labeled information is often a challenging task. Therefore, it is of paramount importance to develop automatic unsupervised semantic segmentation methods to characterize the subsurface targets without labeled information. Unsupervised segmentation methods learn to discover meaningful semantic contents and decompose them into distinct semantic segments with known ontology. Here, we propose an unsupervised radargram segmentation network that uses a convolution-based expansive network as a proxy decoder and a progressive stepwise reconstruction strategy of the input signal from the latent space to measure the spatial similarity with the input radar sounder signal. After designing a unique training strategy by bootstrapping the randomness inside the minibatch and combining the spatial similarity loss along with the contrastive correlation loss, the proposed architecture outperformed the state of the art in fully unsupervised settings. Experiments were conducted on the multichannel coherent radar depth sounder to test the robustness of the proposed method. We carried out a comparative analysis with the state-of-the-art unsupervised and supervised segmentation methods. MIoU is improved by $\text{23.47} \%$.

Keywords

• Multichannel coherent radar depth sounder (MCoRDS)
• radar sounder
• semantic segmentation
• sequence-to-sequence model
• TransFuse
• TransUNet