Autoencoder-based Hyperspectral Anomaly Detection using Kernel Principal Component Pre-Processing

Anomaly detection in hyperspectral remote sensing applications has attracted colossal attention due to its ability to uncover small distinctive objects dispersed across large geographical areas in an unsupervised manner. Autoencoders (AEs) have recently been demonstrated as effective tools for detecting hyperspectral anomalies. Using pre-processing techniques along with AEs improves accuracy by removing noise and irrelevant information from the data and also improves computational efficiency by reducing the dimensionality of the data or transforming it into a more appropriate representation. Therefore, this paper proposes to utilize principal component analysis (PCA) and kernel PCA (KPCA) based pre-processing methods in conjunction with the autonomous hyperspectral anomaly detection autoencoder (AUTO-AD). Further, we propose using multiple kernels in KPCA-based pre-processing to capture the complexity of the data better. Although KPCA- and MKPCA-based pre-processing shows excellent results when combined with hyperspectral anomaly detection algorithms, their high computational cost becomes crucial in resource-constrained applications. As a solution, we use random Fourier features (RFF) to approximate KPCA-based pre-processing. We conduct a series of experiments on various datasets to demonstrate the performance of the proposed framework. The experiments reveal that utilizing KPCA as a pre-processing step lead to better results than state-of-the-art hyperspectral anomaly detection approaches.