CDCTA: cascaded dual-constrained transformer autoencoder for hyperspectral unmixing with endmember variability and spectral geometry

Hyperspectral unmixing (HU) in hyperspectral image (HSI) processing is a crucial step. However, the accuracy of unmixing methods is limited by the variability in endmember and the complexity of the HSI structure found in natural scenes. Endmember variability refers to the variations or differences exhibited by endmembers in different locations or under varying conditions within a hyperspectral remote sensing scene. Therefore, to enhance the accuracy of unmixing results, it is crucial to fully leverage spectral, geometric, and spatial information within HSIs, comprehensively exploring the spectral characteristics of endmembers. We present a cascaded dual-constrained transformer autoencoder (AE) for HU with endmember variability and spectral geometry. The model utilizes a transformer AE network to extract the global spatial features in the HSI. Additionally, it incorporates the minimum distance constraint to account for the geometric information of the HSI. Given the similarity in shape exhibited by endmembers of each individual material, with the primary endmember variability being expressed through overall intensity fluctuations, an abundance-weighted constraint method for endmember spectral angle distance is proposed. During training, the architecture utilizes two cascaded networks to preserve the detailed information in the HSI. We evaluate the proposed model using three real datasets. The experimental results indicate that the proposed method achieves superior performance in abundance estimation and endmember extraction. Furthermore, the effectiveness of the two constraint methods was verified through ablation experiments.