A Comprehensive Overview of Machine Learning Based Feature Extraction Techniques for Hyperspectral Image Classification

Hyperspectral imaging using remote sensing techniques captures important details about the things on Earth by exploiting hundreds of adjacent, tiny spectral bands. The performance is hindered when all the bands are taken into consideration for categorization. Therefore, it is essential to lower the HSI bands, generally by feature selection and extraction. One popular unsupervised feature extraction method is Principal Component Analysis (PCA). Nevertheless, it ignores the local structure of the data in favor of taking into account global variance. Another feature reduction technique Nonnegative Matrix Factorization (NMF), approximates the data in a low-dimensional subspace. The Incremental PCA (IPCA) exploits Singular Value Decomposition (SVD) to transform data to lower space and is suitable for large datasets. Another dimensionality reduction technique Factor Analysis (FA) eradicates band-to-band correlation preserving the vital spectral information. This study investigates the performance among these for feature extraction techniques for effective HSI categorization. The rigorous analysis proves FA as the superior among the other techniques providing an Overall Accuracy of 92.70%, while PCA, NMF and IPCA provide 82.36%, 82.44% and 80.15% respectively.