3D Convolutional Autoencoder with Spectral–Spatial Attention (3DCAE-SSA) for Anomaly Detection in Hyperspectral Imagery

Hyperspectral anomaly detection (HAD) refers to the process of identifying pixels or regions that are significantly different from the surrounding background signature of a hyperspectral image (HSI). HAD is highly challenging due to high spectral dimensionality, complex spatial patterns, subtle nature of the real-world environment, and limited knowledge about the signature of interest. This study develops a HAD algorithm combining three-Dimensional (3D) Convolutional Autoencoder (3DCAE), Spectral-Spatial Attention (SSA) and a robust hybrid loss. We use 3DCAE that employs 3D convolution layers (3DCLs) instead of fully connected layers to create a compact latent space from the input image, while the decoder uses deconvolution layers to recover the input image. We attach multi-headed spectral attention and multi-scale spatial attention that help to expand local receptive fields to global context and allow each voxel to attend to the other long-range voxels. In the proposed 3DCAE-SSA, to minimize the image reconstruction error, we introduce a hybrid robust loss to quantify the reconstruction error instead of using mean squared error (MSE) which is sensitive to outliers. To see the performance of the new algorithm, we build an error map based on a robust statistical Mahalanobis distance. For open access, Urban-4 dataset the AUC values show that the new algorithm performs significantly better having AUC value 99.76% than two existing algorithms: Auto-AD (97.47%), and GAED (99.24%).