A Change Proportion-Driven Small-Sample Learning Framework and an Open-Source Dataset for PolSAR Change Detection

Abstract

Polarimetric synthetic aperture radar (PolSAR) provides rich scattering information to identify subtle or complex changes under all-weather day-and-night conditions. Meanwhile, a series of deep learning (DL) models has brought new breakthroughs to change detection (CD) technology. However, PolSAR CD continues to grapple with issues including the lack of benchmark datasets, varied network inputs, and ever-changing classes of interest. In this article, a dataset containing 2468 co-registered $L$-band UAVSAR image pairs, named WHU-PolSAR-CD, is established to provide a unified benchmark for evaluating the performance of DL-based CD algorithms. It covers various land-cover change scenes, and each image pair is accompanied by carefully curated change masks and rigorous geometric–radiometric consistency checks. Using this dataset, the impact of several representative input types (e.g., real/imaginary parts, linear/logarithmic amplitude and phase, and complex/real-valued elements) on CD performance has been evaluated, showing that the most effective features for discriminating surface changes were logarithmic amplitude and phase, yielding an average F1-score of 90.71%. Furthermore, we propose a novel change proportion-driven small-sample learning (CPSSL) framework for efficiently adapting models to change regions of interest. This framework leverages the global change proportion estimated by the Bartlett distance as weak supervision. CPSSL achieves an F1-score of 77.62% while being trained on only 1% annotated patches, outperforming existing weak and pseudolabel methods, and showing strong resistance to speckle noise.