Heterogeneous Network Based Contrastive Learning Method for PolSAR Land Cover Classification

TL;DR

This work addresses PolSAR land-cover classification under scarce labels and scattering confusion by introducing HCLNet, a heterogeneous-network contrastive learning framework that fuses physical and statistical PolSAR features. It employs a feature filter to reduce feature redundancy and a superpixel-based instance discrimination to generate meaningful positive/negative samples, training a dual-architecture network (online 2-D CNN and target 1-D CNN) with InfoNCE loss. Across three benchmark datasets, HCLNet achieves state-of-the-art performance in both few-shot and full-sample settings, while offering visualization and ablation evidence of each component’s contribution and a favorable model complexity profile. The approach demonstrates the value of unsupervised pretraining on unlabeled PolSAR data for robust, scalable land-cover classification, with implications for leveraging multi-features in PolSAR representation learning.

Abstract

Polarimetric synthetic aperture radar (PolSAR) image interpretation is widely used in various fields. Recently, deep learning has made significant progress in PolSAR image classification. Supervised learning (SL) requires a large amount of labeled PolSAR data with high quality to achieve better performance, however, manually labeled data is insufficient. This causes the SL to fail into overfitting and degrades its generalization performance. Furthermore, the scattering confusion problem is also a significant challenge that attracts more attention. To solve these problems, this article proposes a Heterogeneous Network based Contrastive Learning method(HCLNet). It aims to learn high-level representation from unlabeled PolSAR data for few-shot classification according to multi-features and superpixels. Beyond the conventional CL, HCLNet introduces the heterogeneous architecture for the first time to utilize heterogeneous PolSAR features better. And it develops two easy-to-use plugins to narrow the domain gap between optics and PolSAR, including feature filter and superpixel-based instance discrimination, which the former is used to enhance the complementarity of multi-features, and the latter is used to increase the diversity of negative samples. Experiments demonstrate the superiority of HCLNet on three widely used PolSAR benchmark datasets compared with state-of-the-art methods. Ablation studies also verify the importance of each component. Besides, this work has implications for how to efficiently utilize the multi-features of PolSAR data to learn better high-level representation in CL and how to construct networks suitable for PolSAR data better.

Figures (22)

• Figure 1: Visual comparison of instance similarity between PolSAR and optical images, with PolSAR images on the left and optical images on the right.
• Figure 2: The overall framework of the proposed HCLNet. It mainly contains two processes: Pretraining and Fine-tuning. In pretraining, it first uses Feature Filter to combinate features, then constructs the heterogeneous network and uses Superpixel-based Instance Discrimination to learn the high-level representation. In fine-tuning, it uses the trained online network from pretraining and fine-tunes it with a small number of labeled data to better fit the downstream distribution.
• Figure 3: The architecture of the heterogeneous network in HCLNet. It contains two networks with different architectures and is updated with InfoNCE loss. The output of the target network belonging to different superpixels in the same minibatch will be served as negative samples.
• Figure 4: The architecture of the online network in the heterogeneous network. It contains the representation encoder and the projection head; the former will be used for fine-tuning.
• Figure 5: RADARSAT-2 Flevoland data. (a) Pauli RGB image. (b) Ground-truth image. (c) Superpixel image.