Abstract
So far, a large number of advanced techniques have been developed to enhance and extract the spatially semantic information in hyperspectral image processing and analysis. However, locally semantic change, such as scene composition, relative position between objects, spectral variability caused by illumination, atmospheric effects, and material mixture, has been less frequently investigated in modeling spatial information. Consequently, identifying the same materials from spatially different scenes or positions can be difficult. In this article, we propose a solution to address this issue by locally extracting invariant features from hyperspectral imagery (HSI) in both spatial and frequency domains, using a method called invariant attribute profiles (IAPs). IAPs extract the spatial invariant features by exploiting isotropic filter banks or convolutional kernels on HSI and spatial aggregation techniques (e.g., superpixel segmentation) in the Cartesian coordinate system. Furthermore, they model invariant behaviors (e.g., shift, rotation) by the means of a continuous histogram of oriented gradients constructed in a Fourier polar coordinate. This yields a combinatorial representation of spatial-frequency invariant features with application to HSI classification. Extensive experiments conducted on three promising hyperspectral data sets (Houston2013 and Houston2018) to demonstrate the superiority and effectiveness of the proposed IAP method in comparison with several state-of-the-art profile-related techniques. The codes will be available from the website: https://sites.google.com/view/danfeng-hong/data-code.
Introduction: The Challenge of Local Semantic Variability
Hyperspectral imaging (HSI) captures hundreds of narrow spectral bands, enabling precise material identification. However, complex real-world scenes introduce "local semantic variability"—changes in scene composition, object orientation, and illumination that complicate classification. For instance, two identical structures may appear different to standard algorithms if one is rotated or shifted relative to the other.
Traditional methods like Attribute Profiles (APs) often fail to bridge this "semantic gap" because their spatial fingerprints are sensitive to these variations. As shown in the comparison below, standard APs fluctuate significantly when objects shift or rotate, whereas the proposed Invariant Attribute Profiles (IAPs) maintain a more consistent feature representation.
The IAP Framework
To address these challenges, the authors propose a joint spatial-frequency feature extractor. The IAP framework consists of two primary components:
1. Spatial Invariant Features (SIFs)
SIFs are designed to resist rotation and noise. The process begins by applying isotropic filter banks to the HSI data. Unlike standard directional kernels, isotropic filters are symmetric, making them naturally invariant to rotation. These features are then aggregated using Simple Linear Iterative Clustering (SLIC) superpixel segmentation. By grouping pixels into semantically meaningful regions, the model shifts its focus from individual noisy pixels to stable "objects."
2. Frequency Invariant Features (FIFs)
To handle shifts and rotations more robustly than discrete methods (like standard HOG), the authors utilize the Fourier polar coordinate system. By modeling the rotation of a patch as a continuous signal, they leverage Fourier transform properties to represent rotation as a phase shift. This shift can be mathematically neutralized, resulting in a descriptor that remains constant regardless of the object's orientation.
The resulting high-dimensional vector is condensed using Principal Component Analysis (PCA) before being classified by standard models like Random Forest (RF).
Experimental Results and Insights
The IAP method was evaluated on three major datasets: Pavia University, Houston 2013, and Houston 2018.
- Performance Gains: On the complex Houston 2018 dataset, IAPs achieved an Overall Accuracy (OA) of 83.68%, significantly outperforming baseline spectral features (65.85%) and state-of-the-art methods like Extinction Profiles (70.59%).
- Structural Robustness: IAPs excel at identifying large, regular structures such as highways and non-residential buildings, which are often obscured by shadows or varying textures in traditional models.
- Domain Fusion: The success of IAPs demonstrates that combining spatial information (location) with frequency information (texture) provides a more robust material "signature."
Conclusion
Invariant Attribute Profiles offer a powerful solution for HSI classification by bridging the semantic gap caused by local scene variations. While highly effective for structured objects, future work aims to improve performance on "disorderly" natural classes like vegetation and to integrate these invariance principles into end-to-end deep learning architectures.
Summary updated