Enhancing Hyperspectral Image Classification with Deep Bagging Ensembles

Abstract

Hyperspectral imaging, with its rich spectral information across numerous contiguous spectral bands, provides unprecedented opportunities for precise pixel classification. However, the complexity and high dimensionality of hyperspectral data often pose significant challenges, such as Hughes phenomenon and overfitting, particularly when using deep learning models like Convolutional Neural Networks (CNNs). This research introduces a novel approach to hyperspectral image classification that leverages the robustness of bagging (bootstrap aggregation) integrated with CNNs to enhance classification accuracy and model stability. By implementing a deep bagging ensemble, where multiple CNN models are trained independently on random subsets of the training data and then aggregated, the proposed method aims to reduce model variance without substantially increasing bias. Each CNN acts as a base classifier, operating on bootstrap samples derived from the original dataset, thereby introducing diversity in the model predictions. The ensemble’s final decision is made through majority voting or averaging, which improves generalization over unseen data by mitigating the risk of overfitting to noisy or unrepresentative training samples. Comparative analysis with traditional single-model approaches and other ensemble techniques confirms the efficacy of the deep bagging ensemble in achieving superior classification performance. The method not only addresses the spectral-spatial classification challenges in hyperspectral datasets but also shows significant improvement in stability and accuracy across various challenging datasets. This study's outcomes suggest that deep learning ensembles, particularly those utilizing bagging with CNNs, can effectively tackle the intricacies of hyperspectral image classification, making them suitable for advanced remote sensing applications.