Hypergraph p-Laplacian Regularization for Remote Sensing Image Recognition

Xueqi Ma, Weifeng Liu, Shuying Li, Yicong Zhou

Introduction

Remote sensing image recognition plays a critical role in environmental monitoring, urban planning, and disaster management. However, the field faces a significant challenge: while satellite and aerial imagery are abundant, obtaining high-quality expert annotations is expensive and time-consuming. Semi-supervised learning (SSL) offers a promising solution by leveraging a small amount of labeled data alongside a larger pool of unlabeled data to improve classification performance.

Traditional SSL methods often rely on standard graph Laplacians, which have two primary limitations. First, they typically model only pairwise relationships between data points, failing to capture the complex, high-order structures inherent in remote sensing data. Second, standard graph Laplacians are linear operators ($p=2$), which may not adequately preserve the nonlinear geometry of the data manifold. This paper introduces HpLapR, a novel framework that addresses these gaps by combining hypergraph theory with $p$-Laplacian regularization.

Methodology: The HpLapR Framework

The proposed Hypergraph $p$-Laplacian Regularization (HpLapR) framework integrates high-order relationship modeling with nonlinear manifold regularization. The process begins with feature extraction from both labeled and unlabeled images using a Deep Convolutional Neural Network (CNN).

Overview of the HpLapR framework for remote sensing image classification, illustrating the transition from feature extraction to local structure preservation and final classification.

The core of the methodology consists of three pillars:

  1. Hypergraph Construction: Instead of simple edges connecting two nodes, the authors construct hyperedges that link each image (vertex) with its $k$-nearest neighbors. This allows the model to capture group-level similarities, which is particularly effective for distinguishing between visually similar land-use classes.
  2. The $p$-Laplacian Operator: Moving beyond the standard linear Laplacian, the framework employs a nonlinear $p$-Laplacian. This operator provides greater flexibility in enforcing smoothness across the data manifold, better preserving the underlying geometric structure of the image features.
  3. Efficient Approximation: Since computing exact eigenvectors for a hypergraph $p$-Laplacian is computationally demanding, the authors developed a gradient-descent-based approximation algorithm to ensure the method is scalable for practical applications.

The resulting regularizer is integrated into a Logistic Regression framework. By applying the Representer Theorem, the authors ensure that the optimal classification function can be expressed as a linear combination of kernel functions centered on the training data.

Experimental Setup

The HpLapR method was evaluated using the UC-Merced dataset, a benchmark in the remote sensing community. The dataset contains 2,100 images categorized into 21 distinct land-use classes, such as "agricultural," "airplane," "baseball diamond," and "dense residential." To represent these images, the researchers utilized high-level visual features extracted via a pre-trained CNN, rather than relying on raw pixel values.

Results and Discussion

The researchers compared HpLapR against three baseline methods: standard Laplacian Regularization (LapR), Hypergraph Laplacian Regularization (HLapR), and Graph $p$-Laplacian Regularization (pLapR).

Performance comparison (mAP) of LapR, HLapR, pLapR, and HpLapR across different labeling ratios (30% and 50%).

Key findings from the experiments include:

  • Superior Performance: HpLapR consistently outperformed all baseline methods across various labeling ratios (10%, 20%, 30%, and 50%). The improvement was most pronounced at lower labeling ratios (e.g., 10%), demonstrating the framework's robustness when labeled data is extremely scarce.
  • The Impact of $p$: Validation tests revealed that the standard choice of $p=2$ is often suboptimal. For the UC-Merced dataset, a value of $p=2.6$ yielded the highest mean Average Precision (mAP), confirming the benefits of nonlinear regularization.
  • Handling Class Overlap: The hypergraph approach proved highly effective at distinguishing between overlapping classes, such as "medium density residential" and "dense residential," where local geometric structure is complex.

Conclusion

The HpLapR framework represents a significant advancement in semi-supervised remote sensing image recognition. By unifying hypergraph-based high-order relationship modeling with the nonlinear flexibility of the $p$-Laplacian, the method achieves state-of-the-art results with minimal labeled data. Future work may focus on developing more adaptive hypergraph construction techniques and automating the selection of the optimal $p$ parameter to further enhance performance and scalability.

Summary updated

Abstract

Graph-based and manifold-regularization (MR)-based semisupervised learning, including Laplacian regularization (LapR) and hypergraph LapR (HLapR), have achieved prominent performance in preserving locality and similarity information. However, it is still a great challenge to exactly explore and exploit the local structure of the data distribution. In this paper, we present an efficient and effective approximation algorithm of hypergraph <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${p}$ </tex-math></inline-formula> -Laplacian and then propose hypergraph <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${p}$ </tex-math></inline-formula> -LapR (HpLapR) to preserve the geometry of the probability distribution. In particular, hypergraph is a generalization of a standard graph while hypergraph <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${p}$ </tex-math></inline-formula> -Laplacian is a nonlinear generalization of the standard graph Laplacian. The proposed HpLapR shows great potential to exploit the local structures. We integrate HpLapR with logistic regression for remote sensing image recognition. Experiments on UC-Merced data set demonstrate that the proposed HpLapR has superior performance compared with several popular MR methods including LapR and HLapR.
BibTeX
@article{ma2019hypergraph,
  author = {Xueqi Ma and Weifeng Liu and Shuying Li and Yicong Zhou},
  title = {{Hypergraph p-Laplacian Regularization for Remote Sensing Image Recognition}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2019},
  url = {https://openreview.net/forum?id=4bDiUbLRBk},
}
Figure