GPU Parallel Implementation of Dual-Depth Sparse Probabilistic Latent Semantic Analysis for Hyperspectral Unmixing

Hyperspectral unmixing (HU) is an important task for remotely sensed hyperspectral (HS) data exploitation. It comprises the identification of pure spectral signatures (endmembers) and their corresponding fractional abundances in each pixel of the HS data cube. Several methods have been developed for (semi-) supervised and automatic identification of endmembers and abundances. Recently, the statistical dual-depth sparse probabilistic latent semantic analysis (DEpLSA) method has been developed to tackle the HU problem as a latent topic-based approach in which both endmembers and abundances can be simultaneously estimated according to the semantics encapsulated by the latent topic space. However, statistical models usually lead to computationally demanding algorithms and the computational time of DEpLSA is often too high for practical use, in particular when the dimensionality of the HS data cube is large. In order to mitigate this limitation, this paper resorts to graphical processing units (GPUs) to provide a new parallel version of DEpLSA, developed using the NVidia Compute Device Unified Architecture (CUDA). Our experimental results, conducted using four well-known HS datasets and two different GPU architectures (GTX 1080 and Tesla P100) show that our parallel versions of DEpLSA and the traditional pLSA approach can provide accurate HU results fast enough for practical use, accelerating the corresponding serial versions in at least 30x in the GTX 1080 and up to 147x in the Tesla P100 GPU, which are quite significant acceleration factors that increase with image size, thus allowing for the possibility of fast processing of massive HS data repositories.