Superpixel-guided Discriminative Low-rank Representation of Hyperspectral Images for Classification (2108.11172v2)

Abstract: In this paper, we propose a novel classification scheme for the remotely sensed hyperspectral image (HSI), namely SP-DLRR, by comprehensively exploring its unique characteristics, including the local spatial information and low-rankness. SP-DLRR is mainly composed of two modules, i.e., the classification-guided superpixel segmentation and the discriminative low-rank representation, which are iteratively conducted. Specifically, by utilizing the local spatial information and incorporating the predictions from a typical classifier, the first module segments pixels of an input HSI (or its restoration generated by the second module) into superpixels. According to the resulting superpixels, the pixels of the input HSI are then grouped into clusters and fed into our novel discriminative low-rank representation model with an effective numerical solution. Such a model is capable of increasing the intra-class similarity by suppressing the spectral variations locally while promoting the inter-class discriminability globally, leading to a restored HSI with more discriminative pixels. Experimental results on three benchmark datasets demonstrate the significant superiority of SP-DLRR over state-of-the-art methods, especially for the case with an extremely limited number of training pixels.

• The paper introduces SP-DLRR, a novel hyperspectral image classification method integrating superpixel segmentation and discriminative low-rank representation iteratively.
• SP-DLRR uses local low-rank representation within superpixels and a discriminative model to enhance intra-class similarity and inter-class separation.
• Experiments show SP-DLRR outperforms state-of-the-art methods on benchmark datasets, achieving high accuracy even with limited training data.

The paper "Superpixel-guided Discriminative Low-rank Representation of Hyperspectral Images for Classification" presents a novel approach to hyperspectral image (HSI) classification by leveraging the unique properties of HSIs, specifically their rich spectral and spatial information. The proposed classification scheme, referred to as SP-DLRR, integrates two principal modules: classification-guided superpixel segmentation and discriminative low-rank representation, executed iteratively.

Core Elements:

1. Hyperspectral Characteristics: The authors emphasize the challenges in HSI classification due to spectral variations caused by factors such as illumination and atmospheric conditions. The paper addresses these issues by aiming to recover a robust discriminative representation of HSIs, which is essential for precise classification.
2. Local Low-Rank Representation: SP-DLRR explores the low-rank properties of HSIs at a localized level. Classical methods typically assume global low-rank properties, which can be ineffective for capturing local spatial structures. By applying robust low-rank approximation to superpixels rather than the entire image, the method improves intra-class similarity.
3. Discriminative Representation: The main innovation is incorporating a discriminative low-rank model, which not only enhances intra-class similarity but also promotes inter-class differentiation by reducing spectral variations locally and enhancing discriminability globally. The optimization problem is formulated to minimize local low-rank terms while maximizing inter-class separation using a negative global low-rank term.
4. Iterative Framework: The two modules are applied iteratively: superpixel segmentation refined by classification results, followed by local low-rank representation within these superpixels, to produce a refined HSI representation. This iterative process leads to a progressively more discriminative feature set for classification.
5. Numerical Solution: The optimization problem is tackled using an inexact augmented Lagrange multiplier (IALM) method. Despite the non-convexity introduced by the three-part variable structure, empirical convergence was demonstrated on multiple benchmark datasets.

Experimental Results:

The method outperformed several state-of-the-art HSI classification techniques on benchmark datasets such as Indian Pines, Salinas Valley, and Pavia University. Key advantages include significantly higher overall accuracies (up to 95-99%) and superior performance with limited training samples. The results underline the method's capability in situations with imbalanced class distributions and limited training data, a common scenario in remote sensing applications.

Computational Considerations:

While SP-DLRR offers superior classification performance, it is computationally intensive due to iterative processing and complex low-rank calculations. This trade-off between accuracy and computational demand is noted, suggesting potential areas for efficiency improvements in future works, possibly through advanced optimization acceleration techniques.

In conclusion, the proposed SP-DLRR method provides a robust framework for improving HSI classification by effectively addressing spectral variability and maximizing discriminative power across classes, demonstrating its utility in real-world remote sensing challenges.