Scalable SubXPCA with Extended Morphological Profiles Applied to Hyperspectral Image Classification on a Distributed Platform

The resolution of large volumes of high-dimensional hyperspectral images have increased significantly due to the ongoing advancements in satellite remote sensing technology. One of the most well-known, widely used, and open-source distributed processing framework is Apache Spark, which is proven to quickly handle enormous volumes of these remotely sensed hyperspectral images. Therefore, there is a need for dimensionality reduction algorithms to process these high dimensional images and achieve accurate application performance. This paper suggests a distributed approach for spectral and spatial dimensionality reduction of hyperspectral images. We implemented feature partitioning principal component analysis based on the SubXPCA method for spectral dimensionality reduction and Extended Morphological Profiles for spatial dimensionality reduction in a Spark cluster distributed environment. We compared the proposed method against other feature and non-feature partitioning dimensionality reduction methods with Extended Morphological Profiles. Our experiments on different real and synthetic hyperspectral image datasets confirm that the proposed spectral-spatial reduction method classification performance is better than the spectral reduction methods and their competitors. The proposed method’s running time decreases with an increase in the number of cores in the cluster. The proposed methodology showed a speedup factor of 7.12×, 13.17× and more in a standalone Spark cluster and a Google cloud Dataproc cluster respectively, with an increase in the number of cores.