HSI_Experiments

A light-weight HSI classification framework using custom filtering and Linear SVM

Deep learning has become dominant in hyperspectral image classification owing to its strength in jointly exploiting spectral and spatial information. However, this requires long training times and substantial computational resources. Spectral inconsistency across adjacent bands is a persistent challenge in this domain which CNNs typically address through deep feature learning. This paper proposes a lightweight yet highly effective alternative based on guided image filtering applied across spectral channels prior to classification with a linear SupportVector Machine. The filtering process leverages spatial structure extracted from principal components to perform edge-preserving smoothing, improving spectral consistency while maintaining class boundaries. The proposed method applies a cyclic heterogeneous MNF1-guided filtering across adjacent spectral bands. Adjacent bands share substantial spectral redundancy, so each band benefits from multiple smoothing scales. Despite its simplicity and dataset-agnostic architecture, the proposed Cyclic Heterogeneous MNF1-Guided (CHMF-SVM) framework achieves state-of-the-art performance across seven benchmark datasets i.e. Indian Pines, Pavia Centre, Pavia University, Salinas, SalinasA, Kennedy Space Center (KSC), and Botswana. With only 10% of the labeled data used for training (few-shot), CHMF-SVM attains on average 99.66% OA, 99.03% AA and 99.61% κ, outperforming or matching several recent deep learning, transformer, and state-space models. When the training ratio increases to 30%, the method reaches near-perfect accuracy across all datasets, further highlighting its scalability. The proposed algorithm runs significantly faster (up to 190×) than competing deep learning methods on a standard laptop. The results presented in this paper can be reproduced through the publicly available source codes.