TL;DR
This paper introduces a novel dynamic classifier selection (DCS) approach for hyperspectral image classification that leverages both spectral and spatial information. By considering the neighborhood of a pixel, the method adaptively chooses the best classifier, leading to improved classification accuracy compared to traditional spectral-only methods.
What Problem Does This Paper Address?
Hyperspectral remote sensing is crucial for land use/cover classification, but traditional methods primarily rely on spectral information. While textural features have shown promise, their integration with spectral information in classification, particularly within dynamic classifier selection frameworks, remains an area for improvement. Existing DCS methods often ignore spatial context, which is vital because adjacent pixels are likely to belong to the same class. This paper aims to bridge this gap by developing a DCS approach that dynamically incorporates both spectral and spatial (neighborhood) information to enhance classification accuracy.
Methods
The proposed method introduces a dynamic classifier selection (DCS) algorithm that integrates spectral and spatial information. Key components include:
- Spectral Feature Extraction:
- Principal Component Analysis (PCA): Used for dimensionality reduction of hyperspectral data.
- Minimum Estimated Abundance Covariance (MEAC)-based Band Selection: A supervised method to select the most informative spectral bands, reducing redundancy and improving efficiency.
- Spatial Feature Extraction:
- Volumetric Gray Level Co-Occurrence Matrix (VGLCM): An extension of the traditional GLCM to 3D space, used to extract volumetric textural features from hyperspectral image cubes. This captures relationships between neighboring spectral bands.
- Dynamic Classifier Selection (DCS) with Spectral-Spatial Information:
- A preprocessing step assigns labels to pixels where all member classifiers agree.
- For unlabeled pixels, the algorithm checks the proportion of labeled pixels in their neighborhood.
- If this proportion exceeds a predefined threshold (e.g., 70%), spatial and spectral information are used to determine the label.
- Otherwise, only spectral information is used, employing a DCS-MCB (DCS with Minimum Estimated Abundance Covariance-based band selection) algorithm.
- The VGLCM algorithm is used to extract spatial information.
Datasets
The proposed method was evaluated on two well-known hyperspectral remote sensing datasets:
- HYDICE Washington DC Mall: A subimage scene with 304x301 pixels and 191 spectral bands after bad band removal. It contains six classes: roof, tree, grass, water, road, and trail.
- AVIRIS Indian Pines: A subimage scene with 145x145 pixels and 202 spectral bands. It originally has 16 land-cover classes, but nine classes were used for this experiment.
Key Insights and Ideas
- The Power of Neighborhood: The core idea is that a pixel's label can be more reliably determined by considering its neighbors, especially when those neighbors are already classified. This spatial context significantly aids spectral information.
- Adaptive Classification: The dynamic nature of the classifier selection is crucial. Instead of a one-size-fits-all approach, the algorithm adapts its decision-making process based on the confidence provided by the pixel's neighborhood.
- Synergy of Spectral and Textural Features: Combining spectral information (enhanced by MEAC band selection) with volumetric textural features (from VGLCM) demonstrably improves classification performance beyond using spectral information alone.
- VGLCM for Hyperspectral Texture: Extending GLCM to a 3D volumetric approach (VGLCM) is effective for capturing richer textural information in hyperspectral data, which is often overlooked by 2D methods.
Key Results
- Improved Overall Accuracy: The proposed DCS method achieved improved classification accuracies on both datasets compared to traditional spectral-only methods and a DCS-LCA (Dynamic Classifier Selection - Local Classifier Accuracy) method.
- HYDICE Washington DC Mall: Accuracy improved by 1.13% compared to traditional spectral algorithms.
- AVIRIS Indian Pines: Accuracy improved by 4.47% compared to traditional spectral algorithms.
- Fusion Benefits: Fusion of VGLCM texture features with spectral information (either all bands, PCA components, or selected bands) consistently led to higher accuracies. For example, on the HYDICE dataset, VGLCM 6+all bands achieved an OA of 96.08%.
- Statistical Significance: McNemar's test confirmed that the proposed DCS method provided statistically significant improvements over SVM and DCS-LCA on both datasets.
Limitations and Future Work
- Parameter Sensitivity: While not explicitly detailed as a limitation, the performance of VGLCM is known to be sensitive to parameters like window size and distance. The paper mentions using semivariogram for window size analysis, but further exploration of parameter optimization could be beneficial.
- Computational Cost: Incorporating volumetric texture features and neighborhood analysis can increase computational complexity. The paper doesn't extensively discuss computational efficiency, which is a common trade-off in such methods.
- Future Work: The authors suggest that further research could explore other texture features and advanced ensemble methods to further enhance classification performance. The current work focuses on a specific DCS strategy, and exploring variations or entirely new DCS paradigms could be fruitful.
Summary updated