Randomized Principal Component Analysis for Hyperspectral Image Classification (2403.09117v2)

Abstract: The high-dimensional feature space of the hyperspectral imagery poses major challenges to the processing and analysis of the hyperspectral data sets. In such a case, dimensionality reduction is necessary to decrease the computational complexity. The random projections open up new ways of dimensionality reduction, especially for large data sets. In this paper, the principal component analysis (PCA) and randomized principal component analysis (R-PCA) for the classification of hyperspectral images using support vector machines (SVM) and light gradient boosting machines (LightGBM) have been investigated. In this experimental research, the number of features was reduced to 20 and 30 for classification of two hyperspectral datasets (Indian Pines and Pavia University). The experimental results demonstrated that PCA outperformed R-PCA for SVM for both datasets, but received close accuracy values for LightGBM. The highest classification accuracies were obtained as 0.9925 and 0.9639 by LightGBM with original features for the Pavia University and Indian Pines, respectively.

• The paper evaluates PCA versus R-PCA for hyperspectral image classification using SVM and LightGBM across Indian Pines and Pavia University datasets.
• It finds that PCA generally outperforms R-PCA with SVM, while LightGBM delivers comparable accuracy with both methods when using 30 components.
• The study highlights the need for further research on optimizing dimensionality reduction to balance computational efficiency and classification accuracy.

Comparative Analysis of PCA and R-PCA for Hyperspectral Image Classification with SVM and LightGBM Techniques

Introduction

The classification of hyperspectral images presents a significant challenge due to the high-dimensional feature space inherent in the data. Reducing the dimensionality of these datasets is crucial for efficient processing and analysis. This paper investigates traditional Principal Component Analysis (PCA) and a variant, Randomized Principal Component Analysis (R-PCA), in the context of hyperspectral image classification utilizing Support Vector Machines (SVM) and Light Gradient Boosting Machines (LightGBM). Through an experimental approach, the paper explores the efficacy of these dimensionality reduction methods on Indian Pines and Pavia University hyperspectral datasets, analyzing the impact on classification accuracy of both machine learning algorithms.

Dimensionality Reduction in Hyperspectral Imaging

The necessity of dimensionality reduction in hyperspectral imagery stems from the "curse of dimensionality", which complicates the analysis with conventional techniques. The paper details PCA and R-PCA as projection-based techniques aimed at transforming the high-dimensional data into a lower-dimensional subspace, thus addressing the issue of computational complexity. R-PCA, in particular, is presented as a computationally efficient alternative due to its randomized approach to solving singular value decomposition, suitable for very large datasets.

Classification Methods: SVM and LightGBM

The research employs SVM and LightGBM for hyperspectral image classification, underscoring SVM's prominence in handling high-dimensional data with limited training samples. LightGBM, an advanced ensemble learning algorithm, is noted for its efficiency and effectiveness, attributed to its unique approaches in growing decision trees. This paper is positioned to assess the compatibility of these classifiers with reduced-dimensional features derived from PCA and R-PCA techniques.

Experimental Setup and Results

The experimental design involves the classification of two benchmark hyperspectral datasets, reducing the features to 20 and 30 components using PCA and R-PCA. The classification performance is then evaluated for both SVM and LightGBM classifiers. The highest classification accuracies obtained are 0.9925 for Pavia University and 0.9639 for Indian Pines datasets using LightGBM with original features, indicating a superior performance over the dimensionality-reduced datasets.

In-depth analysis revealed that PCA consistently outperformed R-PCA when coupled with SVM across both datasets. However, for LightGBM, the accuracies between PCA and R-PCA showed minimal differential, particularly with 30 components for the Indian Pines dataset, where R-PCA slightly outperformed PCA. These findings underscore the nuanced impact of dimensionality reduction techniques on classification accuracy, varying by the dataset and classification algorithm employed.

Conclusions and Future Directions

The paper concludes that original features yield better classification results than both PCA and R-PCA, asserting the prominence of LightGBM in achieving the highest classification accuracies. It underscores the limited exploration of R-PCA with machine learning algorithms beyond deep learning for hyperspectral image classification. The paper identifies a research gap, suggesting future investigations into the performance of various classification algorithms with R-PCA-reduced features to fully understand its potential and limitations. This could lead to advancements in computational efficiency and classification accuracy for hyperspectral imagery analysis.

Acknowledgements

Recognition is given to the contributions of Prof. D.A. Landgrebe and Prof. P. Gamba for providing the datasets essential for this research, highlighting the collaborative effort within the scientific community towards advancing remote sensing technologies.

This paper makes a significant contribution to the hyperspectral image analysis field, offering insights into the practical and theoretical implications of PCA and R-PCA in machine learning-based classification tasks. Future research directions will likely focus on optimizing feature reduction techniques and exploring their synergies with various classification algorithms to enhance hyperspectral image analysis further.