Deep Embedding-Based Dynamic Ensemble Selection with Localized Classifier Competence for Hyperspectral Image Classification
Preprint
- 25 March 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Hyperspectral image classification (HSI) faces challenges from high-dimensional redundancy and spatial heterogeneity. Existing methods struggle with computational inefficiency in classifier selection due to diverse architectures and hyperparameters. We propose DEDES, a dynamic ensemble framework that addresses this via spectral-spatial embedding and density-based clustering to partition features into homogeneous regions, enabling localized classifier activation. Experiments on Botswana and Purdue datasets demonstrate state-of-the-art accuracy (overall accuracy of 86.25% and 94.40%) compared to conventional methods. The framework eliminates exhaustive aggregation, demonstrating optimal accuracy through adaptive minimal classifier selection, offering a scalable solution for real-world HSI interpretation.Keywords
This publication has 14 references indexed in Scilit:
- Graph ensemble neural networkInformation Fusion, 2024
- Cross-Domain Distribution Calibration of Hyperspectral Image ClassificationIEEE Geoscience and Remote Sensing Letters, 2023
- Vision Transformer With Contrastive Learning for Hyperspectral Image ClassificationIEEE Geoscience and Remote Sensing Letters, 2023
- Hybrid Transformer-CNN for Real Image DenoisingIEEE Signal Processing Letters, 2022
- Deep learning classifiers for hyperspectral imaging: A reviewISPRS Journal of Photogrammetry and Remote Sensing, 2019
- Deep Learning for Hyperspectral Image Classification: An OverviewIEEE Transactions on Geoscience and Remote Sensing, 2019
- Dynamic ensemble selection for multi-class imbalanced datasetsInformation Sciences, 2018
- Dynamic classifier selection: Recent advances and perspectivesInformation Fusion, 2018
- From dynamic classifier selection to dynamic ensemble selectionPattern Recognition, 2008
- Signal Theory Methods in Multispectral Remote Sensing2003