Spectral-Spatial Feature Extraction of Hyperspectral Images Based on Propagation Filter

Zhikun Chen, Junjun Jiang, Xinwei Jiang, Xiaoping Fang, Zhihua CaiSensors 2018 · DOI

TL;DR

This paper introduces a novel spectral-spatial feature extraction method for hyperspectral images (HSIs) using a Propagation Filter (PF) combined with Principal Component Analysis (PCA). The method effectively addresses the challenge of "cross-region mixing" in HSIs, leading to significantly improved classification accuracy compared to existing techniques.

What Problem Does This Paper Address?

Hyperspectral images (HSIs) are rich in information due to their numerous spectral bands and spatial structures, making them valuable for applications like environmental monitoring and military reconnaissance. However, the high dimensionality of HSIs can lead to the Hughes phenomenon, where classification accuracy decreases with too many features. While dimensionality reduction (DR) and feature extraction are standard preprocessing steps, existing methods often struggle with "cross-region mixing." This occurs when a single pixel, due to limited spatial resolution or complex feature distributions, contains information from multiple distinct regions. This mixing can blur spectral discriminative information and lead to inaccurate classifications. This paper aims to tackle this specific problem by developing a feature extraction method that can effectively suppress the influence of mixed pixels.

Methods

The proposed method, termed PCA-PF-SVM, involves a two-stage feature extraction process followed by Support Vector Machine (SVM) classification.

  1. Principal Component Analysis (PCA): First, PCA is applied to the HSI to reduce its dimensionality. This step aims to remove redundant spectral information and retain the most important features, effectively creating a lower-dimensional representation of the HSI.
  2. Propagation Filter (PF): The core of the novel approach is the Propagation Filter (PF). The PF is a smoothing filter that calculates a new pixel value by considering a weighted sum of its neighboring pixels. The key innovation lies in its weighting mechanism. The filter template assigns weights to neighboring pixels based on their spectral similarity to the central pixel and their spatial proximity. Crucially, when dealing with mixed pixels, the PF's weighting scheme is designed to reduce the influence of pixels from different regions, thereby mitigating the "cross-region mixing" problem. The paper provides detailed equations for calculating these weights, emphasizing how they adapt to suppress the impact of dissimilar neighboring pixels.
  3. Support Vector Machine (SVM): Finally, an SVM classifier is trained on the spectral-spatial features extracted by the PCA-PF method to perform the classification task.

Datasets

The proposed method was evaluated on three widely used hyperspectral image datasets:

  • Indian Pines: Acquired by the AVIRIS sensor, covering agricultural land with 16 ground object categories.
  • Salinas: Acquired by the AVIRIS sensor, covering the Salinas Valley with 16 ground object types.
  • University of Pavia: Acquired by the ROSIS sensor, covering an urban area with multiple ground object types.

The paper details the number of pixels, bands, and spatial resolution for each dataset, along with the number of training and testing samples used in the experiments.

Key Insights and Ideas

  • The "Cross-Region Mixing" Problem is a Significant Hurdle: The paper highlights that even advanced filtering techniques can struggle when pixels are spectrally mixed, leading to degraded classification performance.
  • Propagation Filter's Adaptive Weighting is Key: The PF's ability to dynamically adjust neighbor weights based on spectral and spatial similarity is crucial for suppressing the influence of mixed pixels. This is the central novelty.
  • PCA as a Preprocessing Step Enhances PF: While PCA alone might lose some information, combining it with the PF proves more effective than either method individually. PCA reduces dimensionality and noise, preparing the data for the PF to extract more robust spectral-spatial features.
  • Robustness to Limited Training Data: The proposed method demonstrates strong performance even with a small number of training samples, suggesting its potential for scenarios where labeled data is scarce.

Key Results

The PCA-PF-SVM method consistently outperformed several other state-of-the-art HSI classification methods across the three datasets.

  • Improved Classification Accuracy: The proposed method achieved significant improvements in Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient compared to methods like SVM, PCA-SVM, PCA-Gabor-SVM, and others. For instance, on the Indian Pines dataset, the OA was 13.26% higher than PF-SVM and 25.32% higher than SVM.
  • Effective Cross-Region Mixture Mitigation: Visualizations and quantitative results (especially in areas marked with white boxes in Figure 10) demonstrate the method's ability to better delineate boundaries and reduce misclassifications in regions prone to cross-region mixing.
  • Robustness: The method showed superior performance when the number of training samples per class was varied from 10 to 50, particularly excelling with fewer samples. For example, with 10 training samples per class, the method showed a 3.12–36.31% advantage on the Indian Pines image.
  • Statistical Significance: Paired t-tests confirmed that the observed increases in Kappa values for the PCA-PF-SVM method were statistically significant across all three datasets.

Limitations and Future Work

  • Parameter Sensitivity: The paper acknowledges that the PF's performance is influenced by parameters like the standard deviation of the Gaussian function ($\sigma_a$, $\sigma_r$), the window size ($w$), and the PCA dimension ($k$). While they conducted sensitivity analysis and set optimal parameters for their experiments ($\sigma_a = \sigma_r = 1.5$, $w = 8$, $k = 45$), further research could explore more adaptive or automated parameter selection.
  • Computational Complexity: While not explicitly detailed as a limitation, the combination of PCA and PF filtering, followed by SVM, might incur higher computational costs compared to simpler methods. Future work could focus on optimizing the computational efficiency.
  • Exploration of Other Classifiers: The study primarily uses SVM. Future research could investigate the performance of the PCA-PF feature extraction with other advanced classifiers.
  • Extension to Other Applications: The method's effectiveness in handling spectral-spatial information could be explored for other remote sensing tasks beyond classification.

Summary updated

BibTeX
@article{chen2018spectral,
  author = {Zhikun Chen and Junjun Jiang and Xinwei Jiang and Xiaoping Fang and Zhihua Cai},
  title = {{Spectral-Spatial Feature Extraction of Hyperspectral Images Based on Propagation Filter}},
  journal = {Sensors},
  year = {2018},
  doi = {10.3390/s18061978},
}
Figure