Self-Supervised Learning With Learnable Sparse Contrastive Sampling for Hyperspectral Image Classification

Miaomiao Liang

doi:10.1109/TGRS.2023.3331888

Self-Supervised Learning With Learnable Sparse Contrastive Sampling for Hyperspectral Image Classification

2023-01-01

3

PRE

AI

DOI：10.1109/TGRS.2023.3331888

Original

Original request for help

Save

Abstract

En 中文

Contrastive learning (CL) with learnable examples performs outstandingly in data representation. However, when dealing with hard samples, instance-level alignment with excessive uniformity may descend into trivial clusters, especially when confronted with interclass similarity and intraclass diversity in hyperspectral images (HSIs). To solve this problem, we regard prototypical CL as tracing the potential probability density distribution. Then, a novel pretraining method, learnable sparse contrastive sampling (LSCoSa), is proposed for discriminative representation learning, containing sparse positive sampling and multiple positives learning. Specifically, on the basis of cooperative-adversarial CL, we first exert a Kullback-Leibler (KL) divergence regularizer on the average activation probability of the prototypes, suppressing fake density prototypes for sparse positive sampling. Furthermore, we propose multiple positives learning, in which the top -k potential positives are retrieved and dynamically weighted for contrastive supervision, to avoid trivial clusters and cover satisfying semantic variations. Comprehensive experiments on three HSI benchmark datasets demonstrate that LSCoSa achieves significant advantages over other HSI classification (HSIC) methods. The code is available at https://github.com/sakurashine/LSCoSa.

Keywords:

Contrastive self-supervised learning (SSL)

hyperspectral image (HSI)

Kullback-Leibler (KL) divergence

positive and negative samples

Self-Supervised Learning With Learnable Sparse Contrastive Sampling for Hyperspectral Image Classification

Abstract

Journal

Organization