Superpixel-based two stream spatial spectral fusion autoencoder for hyperspectral target detection

Hyperspectral remote sensing images (HSI), with their integrated spatial–spectral characteristics, offer rich information for target detection. However, the common limitation of having only a single target spectrum in most datasets restricts the effective use of spatial context, which is vital for addressing spectral uncertainty. To overcome this challenge, a superpixel-based patch construction strategy combined with a two-stream spatial–spectral fusion autoencoder (TS 3 FAE) is proposed. A neighborhood spectral patch construction module organizes pixels into spatial–spectral patches reflecting real-world spatial distributions, enabling effective training of target and background representations from limited spectral samples. The model leverages spatial–spectral joint convolution and autoencoding to jointly learn spectral and local spatial features within superpixel neighborhoods. Additionally, a linear background suppression module enhances target–background separability. Experiments on the San Diego, Texas Coast, Bay Champagne, and Cat Island benchmark datasets demonstrate that TS 3 FAE outperforms existing methods. The code for this work will be available at https://github.com/ChevalierZ/TS3FAE for the sake of reproducibility.