Significant improvements in the resolution of enormous high volumes of hyperspectral images in remote sensing technology lead to face challenges in storage, processing, and computing. These challenges are addressed by several open-source distributed processing architectures that have emerged recently for hyperspectral images. This paper proposes a parallel and effective spectral dimensionality reduction approach for classification of hyperspectral images. We designed a novel version of feature partitioning kernel principal component analysis centered on the SubXPCA method applied to the spectral reduction of hyperspectral images in a distributed Spark cluster computing environment. The proposed scalable kernel SubXPCA method is a novel variation of SubXPCA. We compared the scalable kernel SubXPCA method against other feature partitioning spectral dimensionality reduction methods. Our experimental findings on various ground truth and synthetic datasets of hyperspectral image confirm that the proposed spectral reduction approach outperforms its competitors in classification performance for most classifiers. The proposed approach shows increased execution time and speedup as the dataset size grows in the distributed processing environment.