Locality and Structure Regularized Low Rank Representation for Hyperspectral Image Classification (1905.02488v1)

Abstract: Hyperspectral image (HSI) classification, which aims to assign an accurate label for hyperspectral pixels, has drawn great interest in recent years. Although low rank representation (LRR) has been used to classify HSI, its ability to segment each class from the whole HSI data has not been exploited fully yet. LRR has a good capacity to capture the underlying lowdimensional subspaces embedded in original data. However, there are still two drawbacks for LRR. First, LRR does not consider the local geometric structure within data, which makes the local correlation among neighboring data easily ignored. Second, the representation obtained by solving LRR is not discriminative enough to separate different data. In this paper, a novel locality and structure regularized low rank representation (LSLRR) model is proposed for HSI classification. To overcome the above limitations, we present locality constraint criterion (LCC) and structure preserving strategy (SPS) to improve the classical LRR. Specifically, we introduce a new distance metric, which combines both spatial and spectral features, to explore the local similarity of pixels. Thus, the global and local structures of HSI data can be exploited sufficiently. Besides, we propose a structure constraint to make the representation have a near block-diagonal structure. This helps to determine the final classification labels directly. Extensive experiments have been conducted on three popular HSI datasets. And the experimental results demonstrate that the proposed LSLRR outperforms other state-of-the-art methods.

• The paper introduces LSLRR, a novel method enhancing Low Rank Representation for hyperspectral image classification by integrating locality constraints and structure preservation strategies.
• Experimental results on three datasets demonstrate that LSLRR significantly outperforms state-of-the-art methods, achieving substantial improvements in overall accuracy.
• This enhanced approach improves precision for real-world applications and offers a foundation for future research integrating local spatial-spectral structures with advanced machine learning techniques.

Overview of Locality and Structure Regularized Low Rank Representation for Hyperspectral Image Classification

The paper "Locality and Structure Regularized Low Rank Representation for Hyperspectral Image Classification" presents an advanced approach to hyperspectral image (HSI) classification by improving upon the conventional low rank representation (LRR) methods. The novel model, referred to as LSLRR, enhances the original LRR by integrating locality constraint criteria (LCC) and structure preserving strategies (SPS), addressing previously unmet challenges within hyperspectral data classification.

Introduction

Hyperspectral imaging, characterized by acquiring detailed spectral information from across the electromagnetic spectrum, has become a focal point of remote sensing research. The primary aim is to accurately label each pixel within these images. Despite the utility of LRR in identifying low-dimensional subspaces within hyperspectral data, traditional LRR approaches fall short by not adequately capturing local geometric structures and lacking discriminative capabilities necessary for effective classification.

Methodology

The LSLRR model aims to rectify these shortcomings through two key enhancements:

1. Locality Constraint Criterion (LCC): This criterion introduces a new distance metric that blends both spatial and spectral features. It ensures that locality is considered by reinforcing pixel-wise spatial coherence, thus allowing for a more exhaustive exploitation of the intrinsic structures present in hyperspectral data.
2. Structure Preserving Strategy (SPS): LSLRR enforces a block-diagonal structure within the representation matrix. This ensures that pixels from the same class are grouped, thereby improving discriminative power.

Additionally, the model incorporates an adaptive dictionary learning process, allowing the model to formulate a more robust dictionary that moves beyond the constraints of the traditional LRR framework. This is achieved through a discriminative mechanism that enhances classification precision.

Results

Extensive experiments were conducted on three well-regarded hyperspectral datasets: Indian Pines, Pavia University, and Salinas. The results demonstrated that LSLRR consistently outperformed state-of-the-art methods in overall accuracy (OA), average accuracy (AA), and kappa coefficient metrics.

• In the Indian Pines dataset, LSLRR achieved an OA of 95.63%, significantly surpassing the original LRR by 25.16%.
• LSLRR attained an OA of 98.52% on the Pavia University set, which is an improvement of approximately 11.79% over baseline LRR.
• For the Salinas dataset, LSLRR reached an OA of 97.77%, again demonstrating substantial advances over the LRR.

Implications and Future Directions

The contribution of LSLRR is twofold. Practically, it delivers enhanced classification accuracy, making it suitable for real-world applications demanding precision such as environmental monitoring and agricultural assessments. Theoretically, it opens pathways for further exploration into integrating local spatial-spectral structures within other dimensionality reduction frameworks.

Moving forward, there is potential to integrate machine learning models like deep learning with LSLRR, which could refine classification strategies further and address massive datasets' scalability needs. Exploration into real-time applications and incorporating geographical metadata could vastly improve real-world applicability and decision-making processes in various remote sensing fields.