A Lightweight CNN-SEEDS Fusion Model With Factor Analysis for Hyperspectral Image Classification

Abstract

Hyperspectral remote sensing enables detailed characterization of Earth surface materials through hundreds of spectral bands. However, the high dimensionality of hyperspectral images and the limited availability of labeled samples make classification computationally challenging. This work addresses this problem by proposing a lightweight spatial–spectral fusion approach that combines the superpixel-extracted energy-driven sampling (SEEDS) superpixel segmentation algorithm with a 1-D convolutional neural network to improve classification performance while keeping computational cost low. The proposed methodology includes dimensionality reduction of spectral features, spatial segmentation, and classification using a compact neural architecture, while also investigating factor analysis as an alternative to principal component analysis and evaluating different SEEDS parameter configurations. Experiments follow current research practices by considering limited training samples. The results show that the proposed approach achieves competitive classification performance while maintaining reduced architectural complexity and low inference time, highlighting its suitability for resource-constrained environments and for future hardware deployment.