TuckerNet: A Multilayer Unsupervised Hyperspectral Change Detection Framework With Tensor Decomposition

Hyperspectral change detection plays a pivotal role in remote sensing by identifying subtle surface changes through the analysis of bitemporal hyperspectral images (HSIs). Existing unsupervised methods are often limited by sensitivity to pseudo-sample selection, complex hyperparameter tuning, reliance on handcrafted features, and loss of spectral-spatial coupling information. To overcome these limitations, we proposeTuckerNet, a novel multilayer unsupervised change detection framework that leverages 3D tensor representations to fully preserve the inherent spectral-spatial correlations in HSIs. At the core of our framework isTucker Convolution, a convolutional operation built on Tucker decomposition with adaptive rank estimation, enabling efficient extraction of low-dimensional spectral-spatial features without labeled data. Additionally, a novel change detector integrates independent spectral discrepancy and local spatial analysis to achieve adaptive thresholding and generate accurate binary change maps. Extensive experiments across multiple datasets demonstrate thatTuckerNetconsistently outperforms state-of-the-art unsupervised methods, exhibiting strong robustness to hyperparameter variations and highly competitive computational efficiency.