Structured Priors for Sparse-Representation-Based Hyperspectral Image Classification

Pixel-wise classification, where each pixel is assigned to a predefined class, is one of the most important procedures in hyperspectral image (HSI) analysis. By representing a test pixel as a linear combination of a small subset of labeled pixels, a sparse representation classifier (SRC) gives rather plausible results compared with that of traditional classifiers such as the support vector machine (SVM). Recently, by incorporating additional structured sparsity priors, the second generation SRCs have appeared in the literature and are reported to further improve the performance of HSI. These priors are based on exploiting the spatial dependencies between the neighboring pixels, the inherent structure of the dictionary, or both. In this paper, we review and compare several structured priors for sparse-representation-based HSI classification. We also propose a new structured prior called the low rank group prior, which can be considered as a modification of the low rank prior. Furthermore, we will investigate how different structured priors improve the result for the HSI classification.

💡 Research Summary

This paper investigates how to improve hyperspectral image (HSI) classification by enriching the sparse representation classifier (SRC) with various structured sparsity priors. The classic SRC models a test pixel y ∈ ℝ^P as a linear combination of all training pixels (the dictionary A ∈ ℝ^{P×N}) and enforces sparsity on the coefficient vector x via an ℓ₁ norm. While effective in many computer‑vision tasks, SRC suffers on HSI because the dictionary atoms are highly correlated and training samples are often scarce, leading to unstable coefficient estimates.

To address these issues, the authors categorize structured priors into three families: (a) priors that exploit spatial‑spectral dependencies among neighboring pixels, (b) priors that exploit the inherent group structure of the dictionary (each class forms a sub‑dictionary), and (c) priors that combine both aspects. Six specific priors are examined:

1. Joint Sparsity (JS) – assumes a small spatial neighbourhood of T pixels shares exactly the same support. The coefficient matrix X ∈ ℝ^{N×T} is forced to be row‑sparse by minimizing the ℓ₂,₁ norm (∑_i‖x_i‖₂). This yields strong consistency within homogeneous regions but can over‑constrain boundary pixels.

2. Laplacian Sparsity (LS) – introduces a similarity matrix W for the neighbourhood, builds the normalized graph Laplacian L = I − D^{−½}WD^{−½}, and adds a trace term tr(XLXᵀ) to the objective. The resulting regularizer encourages similar pixels to have similar coefficients while allowing dissimilar pixels to diverge, offering more flexibility than JS.

3. Group Lasso (GS) – leverages the fact that the dictionary is naturally partitioned into K class‑wise groups G₁,…,G_K. The regularizer ∑_g w_g‖x_g‖₂ promotes selection of whole groups (i.e., classes) and discarding of the rest. This is especially useful for mixed‑pixel scenarios but does not enforce sparsity inside a selected group.

4. Sparse Group Lasso (SGS) – combines GS with a global ℓ₁ term (λ₁∑_g w_g‖x_g‖₂ + λ₂‖x‖₁). The extra ℓ₁ component induces sparsity within each active group, mitigating the over‑inclusiveness of pure GS. When extended to multiple pixels, the formulation becomes the Collaborative Hierarchical Lasso (CHiLasso).

5. Low‑Rank (LR) Prior – treats the coefficient matrix X as a low‑rank object and penalizes its nuclear norm ‖X‖_* . This encourages the rows of X to lie in a low‑dimensional subspace, which is more tolerant than strict row‑sparsity when neighbourhoods contain mixed classes.

6. Low‑Rank Group (LRG) Prior – the novel contribution of the paper. It merges the low‑rank and group concepts by applying a nuclear‑norm penalty to each class‑wise sub‑matrix X_g and summing them: ∑g w_g‖X_g‖*. Remarkably, a single regularizer simultaneously enforces (i) group‑level sparsity (which classes are active) and (ii) low‑rank structure within each active group (how the coefficients are correlated). This reduces the number of hyper‑parameters compared with CHiLasso while delivering superior performance.

Optimization is performed with standard algorithms: ADMM or SpaRSA for ℓ₁‑type problems, a modified Feature‑Sign Search (FSS) for Laplacian regularization, and Singular Value Thresholding for nuclear‑norm terms. The authors evaluate all priors on two real HSI datasets—Indian Pine (AVIRIS, 145 × 145 pixels, 200 bands after removing noisy bands, 16 classes) and University of Pavia (ROSIS, 610 × 340 pixels, 103 bands, 9 classes)—as well as a toy example that visualizes the sparsity patterns each prior produces.

Training dictionaries are built from a random subset of pixels (≈10 % of labeled samples for Indian Pine, ≈2 % for Pavia). The remaining pixels serve as test data. Performance metrics include Overall Accuracy (OA), Average Accuracy (AA), and the Kappa coefficient (κ).

Key Findings

Indian Pine: The Low‑Rank Group prior achieves the highest OA (92.58 %) and κ (0.923). Laplacian sparsity follows with OA ≈ 84 %, while Joint Sparsity lags at ≈ 71 %. Low‑Rank alone reaches OA ≈ 86 %, confirming that incorporating rank information helps, but the combined LRG yields the best balance of inter‑class discrimination and intra‑class coherence. Visual maps show LRG sharply delineates class boundaries and is robust to noise.

University of Pavia: Again LRG leads with OA ≈ 81 % and κ ≈ 0.843. Group Lasso and Low‑Rank are close (≈ 80 % OA) but still below LRG. Joint Sparsity and Laplacian sparsity perform considerably worse (OA ≈ 66–74 %). The results demonstrate that the joint exploitation of dictionary grouping and low‑rank structure is beneficial across datasets with different spatial resolutions and class distributions.

Class‑wise analysis reveals that classes with very few training samples (e.g., class 7 in Indian Pine) remain challenging for all methods, underscoring the need for additional strategies when training data are extremely scarce. Nonetheless, LRG consistently provides the most stable performance across all classes.

Implications

The study confirms that structured sparsity priors can substantially improve SRC‑based HSI classification by embedding spatial context and dictionary organization directly into the optimization. The proposed Low‑Rank Group prior is particularly attractive because it requires only one regularization term yet captures two complementary forms of structure, simplifying hyper‑parameter tuning and reducing computational overhead relative to multi‑term models like CHiLasso.

Limitations and Future Work

The nuclear‑norm operations are computationally intensive for large dictionaries, suggesting a need for more efficient or incremental low‑rank solvers. Parameter selection (λ, group weights w_g) still relies on cross‑validation; automated or adaptive schemes would enhance practicality. Moreover, the experiments use relatively small training sets; extending the approach to semi‑supervised or active‑learning frameworks could further mitigate data scarcity. The authors propose integrating deep learning for dictionary learning, developing online low‑rank algorithms for real‑time processing, and exploring multi‑scale neighbourhood definitions to better handle heterogeneous regions.

In summary, this paper provides a thorough comparative analysis of structured sparsity priors for SRC in hyperspectral image classification and introduces a novel low‑rank group prior that achieves state‑of‑the‑art accuracy with reduced model complexity, offering a solid foundation for future advances in high‑dimensional remote sensing classification.