Abstract
In this paper, we present a kernel sparse subspace clustering with spatial max pooling operation (KSSC-SMP) algorithm for hyperspectral remote sensing imagery. Firstly, the feature points are mapped from the original space into a higher dimensional space with a kernel strategy. In particular, the sparse subspace clustering (SSC) model is extended to nonlinear manifolds, which can better explore the complex nonlinear structure of hyperspectral images (HSIs) and obtain a much more accurate representation coefficient matrix. Secondly, through the spatial max pooling operation, the spatial contextual information is integrated to obtain a smoother clustering result. Through experiments, it is verified that the KSSC-SMP algorithm is a competitive clustering method for HSIs and outperforms the state-of-the-art clustering methods.
Introduction
Hyperspectral images (HSIs) capture hundreds of narrow spectral bands, providing a unique "fingerprint" for every pixel. However, clustering these pixels into meaningful categories like crops, soil, or urban materials is challenging due to two primary factors:
- Nonlinearity: Land-cover materials often exhibit complex, nonlinear relationships that traditional linear models fail to capture.
- Spectral Variability: Environmental factors like lighting and shadows cause the same material to appear different, while distinct materials may appear spectrally similar.
Standard Sparse Subspace Clustering (SSC) assumes that pixels from the same class reside in a linear subspace. In practice, HSIs are often "linearly inseparable." Furthermore, many algorithms treat pixels independently, ignoring the spatial context where neighboring pixels likely belong to the same class.
Methodology
The proposed Kernel Sparse Subspace Clustering with Spatial Max Pooling (KSSC-SMP) algorithm addresses these challenges through a four-stage workflow:
- Kernel Mapping: Data is projected into a higher-dimensional Hilbert space using a Radial Basis Function (RBF) kernel. This "lifting" makes previously nonlinear data linearly separable.
- Sparse Representation: The algorithm employs the Alternating Direction Method of Multipliers (ADMM) to find a sparse coefficient matrix, describing each pixel as a combination of others.
- Spatial Max Pooling: Instead of raw coefficients, the algorithm applies a max pooling operation over a local window (e.g., $3 \times 3$). This suppresses noise and enforces spatial consistency, leading to smoother results compared to standard averaging.
- Spectral Clustering: A similarity graph is constructed from the pooled coefficients, and spectral clustering is applied to generate final labels.
Experimental Results
The algorithm was evaluated on two benchmark datasets: AVIRIS Indian Pines (agricultural) and ROSIS Pavia University (urban).
Quantitative Performance
KSSC-SMP demonstrated significant improvements in Overall Accuracy (OA):
- Indian Pines: Achieved an OA of 80.37%, a substantial increase over the 67.25% achieved by standard SSC.
- Pavia University: Achieved an OA of 74.63%, outperforming specialized methods like $S^4C$ (73.59%).
Qualitative Analysis
Visually, KSSC-SMP produces "cleaner" cluster maps. While traditional methods often suffer from "salt-and-pepper" noise (misclassified isolated pixels), KSSC-SMP generates homogeneous regions that align closely with ground truth. The synergy between the kernel strategy and spatial pooling is evident: while KSSC (kernel only) and SSC-SMP (spatial only) improve upon basic SSC, their combination in KSSC-SMP yields the best results.
Conclusion
KSSC-SMP is a robust method for hyperspectral image clustering, effectively leveraging nonlinear feature mapping and spatial context. Despite its success, the algorithm is sensitive to parameters like the regularization tradeoff ($\lambda$) and kernel parameter ($\delta$). Future research will focus on adaptive parameter tuning and the extraction of more discriminative spatial features to further enhance clustering boundaries.
Summary updated