Abstract:Supervised deep neural networks (DNNs) have proven to be powerful tools for synthetic aperture radar (SAR) image interpretation tasks. However, they present a formidable challenge in acquiring a substantial amount of labeled data. In this article, we investigate the promising technique of contrastive self-supervised learning for SAR image classification. This approach allows us to take advantage of a large number of available unlabeled images to pretrain an SAR image classification model. Our novel contrastive learning framework conducts both instance-level and cluster-level pretext tasks, which not only enforce consistency between the images and their augmented “views” at the instance level but also their representation within clusters. Besides generating different views through random, low-level image transformations, we proposed two new strategies to construct positive sample pairs to improve contrastive SAR image feature learning: middle-level optical assistance and high-level graph searching. The middle-level optical assistance strategy is inspired by the observation that domain experts typically interpret SAR images with the aid of optical images. This insight spurs us to generate intermediate SAR images as positive samples from geographically matched optical data using CycleGAN. Furthermore, we augment the positive samples of the image with their $k$ -nearest neighbor (KNN) counterparts, following the idea that the KNN samples should belong to the same cluster. Extensive experimental results conducted on the SEN12MS land cover classification benchmark dataset demonstrate that our method is competitive with state-of-the-art self-supervised methods for SAR image classification. Even with only a small amount of labeled data for fine-tuning the model, our method rapidly improves classification performance, surpassing models pretrained on natural image datasets.

A Novel Locally Active Learning Method for SAR Image Classification.

Active Learning SAR Image Classification Method Crossing Different Imaging Platforms

A Novel Semi-Supervised Method Based on Active Sample Proposal and Modified Self-Learning for SAR Target Discrimination

Novel Batch Active Learning Approach and Its Application to Synthetic Aperture Radar Datasets

Graph-based Active Learning for Semi-supervised Classification of SAR Data

Sar target classification with limited data via data driven active learning

Simulation-assisted SAR Target Classification Based on Unsupervised Domain Adaptation and Model Interpretability Analysis

An Active Deep Learning Approach for Minimally-Supervised Polsar Image Classification

Multichannel semi-supervised active learning for PolSAR image classification

Learning from Label Proportions for SAR Image Classification

Active Learning in the Spatial Domain for Remote Sensing Image Classification

Remote Sensing Image Classification with Few Labeled Data Using Semisupervised Learning

SAR Image Classification Using Fully Connected Conditional Random Fields Combined with Deep Learning and Superpixel Boundary Constraint

A Novel Semi-Supervised Convolutional Neural Network Method for Synthetic Aperture Radar Image Recognition

LocSAR: a location-based SAR ship generation model for complex backgrounds

Discovering novel categories in sar images in open set conditions

Self-Supervised Classification of SAR Images with Optical Image Assistance

Locality-constraint discriminant feature learning for high-resolution SAR image classification.

A Bayesian Approach to Active Self-Paced Deep Learning for SAR Automatic Target Recognition

Classification of Large-Scale High-Resolution SAR Images with Deep Transfer Learning

A Three-Stage Framework with Reliable Sample Pool for Long-Tailed Classification.