A Comparative Framework of Stacking, Bagging, and Boosting Ensembles for Deep Learning-Based Hyperspectral Image Classification

Abstract

In the realm of remote sensing, hyperspectral image (HSI) classification serves as a pivotal technique for interpreting the vast information conveyed by the electromagnetic spectrum captured in these images. This study delves into the comparative effectiveness of three prominent ensemble learning techniques: Stacking, Bagging, and Boosting, specifically tailored for deep learning-based HSI classification. The research harnesses the diverse landscapes of the Indian Pines, Pavia University, and Salinas datasets to benchmark the performance of these ensemble methods. The stacking ensemble in this study combines Multi-Layer Perceptrons (MLP), Support Vector Machines (SVM), and Convolutional Neural Networks (CNN) with a meta-classifier that integrates the individual predictions into a final decision, aiming to leverage the strengths of different learning models. In contrast, the bagging approach employs multiple CNN models to promote model variance reduction by averaging results, thus improving the robustness of the classification. Meanwhile, the boosting ensemble utilizes Adaptive CNNs that sequentially focus on difficult-to-classify instances, enhancing classification accuracy progressively. An ablation study forms a core component of this research, providing insights into how each ensemble strategy impacts the overall classification performance. This study meticulously evaluates the accuracy, precision, and recall metrics to determine the optimal ensemble approach for HSI classification.