Abstract
This study explores advanced approaches for the segmentation of hyperspectral images by integrating spectral feature systems—including spectral curve, transformation, and metric features—with probabilistic graphical models. We propose a Decomposable Markov Network (DMN) based on the Spectral Angle (SA) metric. By modeling the statistical dependencies between spatial and spectral characteristics, the DMN provides an optimized framework for image segmentation, effectively functioning as a probabilistic neural network. Experimental results demonstrate that the SA-based DMN successfully synthesizes spatial and spectral information, offering superior performance over traditional manual thresholding methods.
Introduction
Hyperspectral imaging provides a wealth of information by capturing data across numerous narrow spectral bands. However, this high dimensionality introduces significant challenges, often referred to as the "dimensionality bottleneck." Key issues include high data redundancy, the difficulty of balancing fine spectral detail with spatial consistency, and the reliance on manual thresholding, which is often subjective and inconsistent across varying environmental conditions. This paper presents an automated segmentation approach using Decomposable Markov Networks (DMN) to address these limitations.
Methodology
Spectral Angle (SA) Metric
To ensure robustness against variations in light intensity and shadows, the proposed method utilizes the Spectral Angle (SA) rather than simple brightness values. The SA measures the "shape" of the spectral curve by calculating the vector angle between spectra. This focus on spectral shape allows for more consistent identification of materials regardless of illumination changes.
Decomposable Markov Networks (DMN)
The core of the segmentation process is the DMN, a probabilistic graphical model. The system calculates the Spectral Angle Cosine (SAC) between adjacent pixels to model local dependencies. A network is constructed where nodes represent probability levels of SAC values, and edges represent transition probabilities. This "gray box" approach provides a balance between the interpretability of statistical models and the automation of neural networks.
Optimization and Pathfinding
The segmentation is treated as an optimization problem. The Floyd-Warshall algorithm is employed to find the Maximum A Posteriori (MAP) path within the DMN. This process allows the network to automatically determine the most likely boundaries between different land cover types based on global statistical evidence, effectively acting as an automated thresholding engine.
Experimental Results and Discussion
The method was evaluated using airborne hyperspectral data from the Yan'an region in China, featuring 63 spectral bands. The study area included complex terrains such as mountains, farmland, rivers, and small structures.
The DMN-based approach was compared against manual thresholding techniques using a comprehensive evaluation score:
- DMN-based Segmentation: 93.66
- Manual Thresholding (0.99): 91.56
- Manual Thresholding (0.991): 90.80
The DMN method demonstrated significant advantages in maintaining spatial cohesion while preserving fine details. Notably, it successfully identified small farmhouses hidden in shadows within dense vegetation—features that were invisible in standard color composites. Unlike manual methods that often lead to over-segmentation, the DMN approach produced clearer, more logical boundaries for linear features like roads and rivers.
Conclusion
The integration of Spectral Angle metrics with Decomposable Markov Networks provides a robust and automated solution for hyperspectral image segmentation. By leveraging the statistical relationships between neighboring pixels, the DMN method overcomes the limitations of manual thresholding and improves resilience to lighting variations. Future research will focus on incorporating higher-order neighbor relationships to further reduce fragmentation and exploring hybrid optimization criteria such as Minimum Description Length (MDL) to refine network generation.
Summary updated