Deep hyperspectral clustering using attention-enhanced 3D-2D convolutional autoencoder for mineral mapping

Remotely sensed hyperspectral images (HSIs) provide valuable compositional information on the surface target materials crucial for Earth observation and geoscience applications. The necessity of labelled data and the high dimensionality of HSI hinder efficient hyperspectral data processing. Hyperspectral data clustering can help to address this challenge. Conventional clustering approaches mainly use shallow spectral absorption features. Deep-learning-based methods, such as autoencoder models, can extract deep HSI's spectral and spatial features. However, the most commonly used 3D-convolutional autoencoder (3D-CAE) models have several disadvantages, including intensive computational costs and the potential to lose spatial information. To avoid losing important information and reduce computational costs, this research proposes an attention-enhanced hybrid 3D-2D-CAE spectral-spatial model for clustering HSIs in mineral mapping. The proposed model enables the capture of non-linear relationships between data points in an unsupervised manner. The network utilizes spectral and spatial attention layers in the 3D and 2D convolutions to capture spectral and spatial information, reducing spectral complexities and enhancing spatial features. The captured feature representations are fed to an agglomerative Gaussian mixture model (AGMM) to cluster HSI. Experimenting with different autoencoder-based clustering methods and comparing their results with those of conventional clustering algorithms, the proposed model achieved an overall accuracy of 88.14 %. It consistently demonstrated the superior performance of the hybrid attention-enhanced 3D-2D-CAE-based clustering method, reinforcing its potential for accurate mineral mapping. Furthermore, the less computationally expensive, attention-enhanced 3D-2D-CAE structure outperforms the extensive GPU usage and processing time of the 3D-CAE method.