Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PCM-SAR: Physics-Driven Contrastive Mutual Learning for SAR Classification

Pengfei Wang, Hao Zheng, Zhigang Hu, Aikun Xu, Meiguang Zheng, Liu Yang

TL;DR

The study tackles SAR image classification under limited labeled data by introducing PCM-SAR, a physics-driven contrastive learning framework. It combines a GLCM-informed noise generation (NSG) and semantic-aware local sampling (SSG) with a multi-level feature fusion using mutual learning to refine feature representations, formalized through loss terms $\\mathcal{L}_1$ and $\\mathcal{L}_2$. key components include $F=\\alpha \\cdot \\log(1+\\beta_1 C) + \\gamma \\cdot \\sqrt{\\beta_2 E} + \\delta \\cdot \\exp(\\beta_3 H)$ and $\\tilde{x}= x \\cdot v'_a \\cdot e^{v_p}$ with $v'_a \\sim \\mathcal{N}(1, (\\sigma')^2)$, as well as $\\hat{z}_{b2}=\\mathrm{softmax}( (W^{Q} f_{1})(W^{K} f_{4})^T / \\sqrt{d_k} ) \\cdot (W^{V} f_{cat})$. Experiments on OpenSARShip and FUSAR-Ship show PCM-SAR surpasses state-of-the-art CL methods and improves small-model performance, including effective transfer to semi-supervised and dense prediction tasks, demonstrating practical gains in SAR analysis.

Abstract

Existing SAR image classification methods based on Contrastive Learning often rely on sample generation strategies designed for optical images, failing to capture the distinct semantic and physical characteristics of SAR data. To address this, we propose Physics-Driven Contrastive Mutual Learning for SAR Classification (PCM-SAR), which incorporates domain-specific physical insights to improve sample generation and feature extraction. PCM-SAR utilizes the gray-level co-occurrence matrix (GLCM) to simulate realistic noise patterns and applies semantic detection for unsupervised local sampling, ensuring generated samples accurately reflect SAR imaging properties. Additionally, a multi-level feature fusion mechanism based on mutual learning enables collaborative refinement of feature representations. Notably, PCM-SAR significantly enhances smaller models by refining SAR feature representations, compensating for their limited capacity. Experimental results show that PCM-SAR consistently outperforms SOTA methods across diverse datasets and SAR classification tasks.

PCM-SAR: Physics-Driven Contrastive Mutual Learning for SAR Classification

TL;DR

The study tackles SAR image classification under limited labeled data by introducing PCM-SAR, a physics-driven contrastive learning framework. It combines a GLCM-informed noise generation (NSG) and semantic-aware local sampling (SSG) with a multi-level feature fusion using mutual learning to refine feature representations, formalized through loss terms and . key components include and with , as well as . Experiments on OpenSARShip and FUSAR-Ship show PCM-SAR surpasses state-of-the-art CL methods and improves small-model performance, including effective transfer to semi-supervised and dense prediction tasks, demonstrating practical gains in SAR analysis.

Abstract

Existing SAR image classification methods based on Contrastive Learning often rely on sample generation strategies designed for optical images, failing to capture the distinct semantic and physical characteristics of SAR data. To address this, we propose Physics-Driven Contrastive Mutual Learning for SAR Classification (PCM-SAR), which incorporates domain-specific physical insights to improve sample generation and feature extraction. PCM-SAR utilizes the gray-level co-occurrence matrix (GLCM) to simulate realistic noise patterns and applies semantic detection for unsupervised local sampling, ensuring generated samples accurately reflect SAR imaging properties. Additionally, a multi-level feature fusion mechanism based on mutual learning enables collaborative refinement of feature representations. Notably, PCM-SAR significantly enhances smaller models by refining SAR feature representations, compensating for their limited capacity. Experimental results show that PCM-SAR consistently outperforms SOTA methods across diverse datasets and SAR classification tasks.

Paper Structure

This paper contains 17 sections, 13 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. METHODOLOGY
  3. Sample Generation Driven by Physical Mechanism
  4. Noise Sample Generation(NSG) based on GLCM
  5. Semantic-aware Sample Generation(SSG)
  6. Mutil-level Feature Fusion(FF) Based on Mutual Learning
  7. Feature Fusion Process
  8. Loss Calculation
  9. EXPERIMENT AND RESULT ANALYSIS
  10. Compared with traditional GLCM methods and ordinary speckle noise sample generation schemes
  11. Comparison with the SOTA CL Methods
  12. Effect of hyper-parameters $\lambda _{1}$ and $\lambda _{2}$
  13. Ablation experiments
  14. Transfer to Semi-Supervised Learning Scenarios
  15. Transfer to Detection and Segmentation Tasks
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Overview of PCM-SAR. (a) is the GMCL nosie noise sample generation process, (b) is the sematic-aware sample generation process. Each model contains an encoder and a momentum encoder. The momentum encoder does not accept gradients but uses EMA(Exponential Moving Average) to update. The encoder accepts two types of losses: contrastive learning loss $L_{sm}$ and mutual learning loss $L_{ml}$.
  • Figure 2: The images represent a comparison of five noise processing effects: (1) is original image, (2) is GLCM-based speckle noise, (3) is regular speckle noise, (4) is Gaussian noise(Commonly used in optical images), and (5) is scattering masking noise, (6) is time-shift noise. Among these, (3), (4), and (6) all significantly disrupt the semantic information of the original image, (5) will cover up the real texture and edge features in the image, which can negatively impact model training.
  • Figure 3: The feature fusion process inside the projection head.