Dual-branch PolSAR Image Classification Based on GraphMAE and Local Feature Extraction

TL;DR

The paper addresses label-scarce PolSAR image classification by proposing a dual-branch architecture that combines a GraphMAE-based superpixel branch with a CNN-based pixel branch, followed by feature fusion for final predictions. The superpixel branch learns region-level polarimetric representations through masked node reconstruction on a superpixel graph using a four-layer Graph Attention Network encoder and a one-layer decoder with a cosine-based loss $L_{SCE}$ and $gamma$, while the pixel branch captures fine-grained pixel features via a four-layer CNN. Experiments on the Flevoland dataset show that the dual-branch model substantially improves performance over single-branch baselines, achieving higher OA and AA and ensuring most classes exceed 94% accuracy, demonstrating effective learning under limited labels. The work highlights the value of generative self-supervised learning for PolSAR and points toward future integration of complex-valued architectures to fully exploit polarimetric information.

Abstract

The annotation of polarimetric synthetic aperture radar (PolSAR) images is a labor-intensive and time-consuming process. Therefore, classifying PolSAR images with limited labels is a challenging task in remote sensing domain. In recent years, self-supervised learning approaches have proven effective in PolSAR image classification with sparse labels. However, we observe a lack of research on generative selfsupervised learning in the studied task. Motivated by this, we propose a dual-branch classification model based on generative self-supervised learning in this paper. The first branch is a superpixel-branch, which learns superpixel-level polarimetric representations using a generative self-supervised graph masked autoencoder. To acquire finer classification results, a convolutional neural networks-based pixel-branch is further incorporated to learn pixel-level features. Classification with fused dual-branch features is finally performed to obtain the predictions. Experimental results on the benchmark Flevoland dataset demonstrate that our approach yields promising classification results.

Figures (1)

• Figure 2: Experimental data and results. (a) PauliRGB image. (b) Ground-truth. (c) CNN. (d) GNN. (e) DB-GC.