Dimensionality Reduction based Convolutional Network for Hyperspectral Target Detection

Hyperspectral target detection relies on exploiting rich spectral-spatial information, yet many deep learning approaches suffer from excessive architectural complexity and high computational cost. To address this issue, we propose a lightweight hybrid 3D-2D convolutional neural network that achieves strong detection performance with significantly reduced computational overhead. The network employs 3D convolutions to capture local spectral-spatial interactions, followed by 2D convolutions for efficient spatial refinement, resulting in a compact architecture with fast inference. To further improve efficiency, unsupervised dimensionality reduction techniques—including Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP)—are applied prior to classification. The proposed framework is evaluated on four benchmark hyperspectral datasets, demonstrating a favorable balance between detection accuracy and processing time. Experimental results indicate that the method achieves competitive AUC performance while requiring substantially lower c