A Hyperspectral Image Change Detection Framework With Self-Supervised Contrastive Learning Pretrained Model

Abstract

Hyperspectral images (HSIs) have high spatial resolution and spectral resolution, and using HSI as a change detection (CD) data source is crucial for detecting surface changes. However, there is a large amount of real noise in HSIs, and most deep learning-based CD methods require a large number of ground-truth labels for training, which is difficult and expensive to label manually. To reduce the dependence of CD on ground-truth labels and weaken the interference of noise on CD in HSIs, in this paper we propose a hyperspectral image change detection framework with self-supervised contrastive learning pre-trained model (CDSCL). CDSCL consists of two parts: self-supervised contrastive learning pre-trained model and CD classification network. The main contributions of this article are as follows: 1) a data augmentation strategy based on Gaussian noise is proposed to improve the ability of the model to extract variation information from HSIs with different random Gaussian noises; 2) based on Information Bottleneck (IB) theory, a progressive feature extraction module (PFEM) is developed to remove redundant or irrelevant details in changing information spectrum; 3) a contrastive loss function based on Pearson correlation coefficient and negative cosine correlation is designed to make the features extracted by the two branches of the siamese network close to each other. Experimental results on four real hyperspectral datasets demonstrate that the CD performance of CDSCL outperforms the most representative CD methods.

Semi-Supervised HyperMatch-Driven Cross Temporal and Spatial Interaction Transformer for Hyperspectral Change Detection

“…Additionally, TFR-PS 2 ANet (2023), which incorporates an attention mechanism-based module from previous work, is included in our comparisons as well [2]. To show the effectiveness of HyperMatch, the pseudo labeling pre-text task is applied for comparison [48]. We use the default parameter settings introduced in the corresponding works for these deep learning-based methods.…”

“…In other words, the labeling cost for training samples is too expensive for hyperspectral analysis [20], especially in hyperspectral CD. There have been a few attempts to address this challenge through active learning approaches [21], [22] or pseudo label strategy [48], they still face limitations due to the inevitable requirement of human participation. Consequently, these limitations continue to impact the performance of current deep learning models, restraining their ability to accurately detect changes in realworld scenarios.…”

Contrastive Self-Supervised Two-Domain Residual Attention Network with Random Augmentation Pool for Hyperspectral Change Detection

“…Ou et al have introduced a self-supervised CL framework for hyperspectral change detection. With the Gaussian noise data augmentation, the feature extraction model is pre-trained using a contrastive loss function [23]. Moreover, based on similarity metrics, the contrastive loss function is designed to strictly recognize the positive sample from a large number of negative samples.…”

“…To address these problems, the random data-augmentation pool is proposed to generate an augmented version of training samples. In this article, random Gaussian noise [23,35], the rotate operation, and the flip operation are adopted to simulate these real scenarios.…”

“…In existing deep hyperspectral change-detection methods, the reality of timeconsuming labeling and unreliable visual interpretation is often ignored. To address this problem, Ou et al introduce self-supervised contrastive learning, which is one type of unsupervised learning method, into hyperspectral change detection with random Gaussian noise to obtain a pre-trained model [23]. However, the downstream change-detection task highly relies on both an upstream contrastive training strategy and a sophisticated model design.…”