Naive Gabor Networks for Hyperspectral Image Classification

Chenying Liu, Jun Li, Lin He, Antonio Plaza, Shutao Li, Bo Li

Abstract

Recently, many convolutional neural network (CNN) methods have been designed for hyperspectral image (HSI) classification since CNNs are able to produce good representations of data, which greatly benefits from a huge number of parameters. However, solving such a high-dimensional optimization problem often requires a large amount of training samples in order to avoid overfitting. Additionally, it is a typical non-convex problem affected by many local minima and flat regions. To address these problems, in this paper, we introduce naive Gabor Networks or Gabor-Nets which, for the first time in the literature, design and learn CNN kernels strictly in the form of Gabor filters, aiming to reduce the number of involved parameters and constrain the solution space, and hence improve the performances of CNNs. Specifically, we develop an innovative phase-induced Gabor kernel, which is trickily designed to perform the Gabor feature learning via a linear combination of local low-frequency and high-frequency components of data controlled by the kernel phase. With the phase-induced Gabor kernel, the proposed Gabor-Nets gains the ability to automatically adapt to the local harmonic characteristics of the HSI data and thus yields more representative harmonic features. Also, this kernel can fulfill the traditional complex-valued Gabor filtering in a real-valued manner, hence making Gabor-Nets easily perform in a usual CNN thread. We evaluated our newly developed Gabor-Nets on three well-known HSIs, suggesting that our proposed Gabor-Nets can significantly improve the performance of CNNs, particularly with a small training set.

Introduction

Hyperspectral images (HSI) provide a rich "spectral fingerprint" for every pixel by capturing hundreds of narrow, contiguous spectral bands. While Convolutional Neural Networks (CNNs) have become the standard for classifying such complex data, they face significant challenges in the HSI domain. Primarily, the "Small Sample Problem" arises because standard CNNs possess millions of free parameters, requiring massive labeled datasets to avoid overfitting—data that is often prohibitively expensive to collect in remote sensing. Furthermore, the unconstrained nature of standard kernels often leads to the learning of redundant or noisy features, complicating the optimization process.

To mitigate these issues, this research introduces Gabor-Nets. By baking a priori knowledge into the architecture, the network constrains its kernels to the mathematical form of Gabor filters. These filters mimic the human primary visual cortex's perception of texture and orientation, providing a structured functional form that regularizes the learning process.

Methodology: The Phase-Induced Gabor Kernel

The core innovation of Gabor-Nets is the Phase-induced Gabor kernel. Unlike traditional Gabor filtering, which is complex-valued and difficult to integrate into real-valued CNN frameworks, the authors propose a real-valued formulation:

$$G(x, y) = K \cos(M + P)$$

Where $K$ is a Gaussian envelope, $M$ represents the harmonic component, and $P$ is the learnable phase offset. This approach reduces the parameter count drastically; instead of learning every weight in a $k \times k$ kernel, the network only learns four parameters: orientation ($\theta$), frequency ($\omega$), scale ($\sigma$), and phase ($P$).

The Gabor convolution process, illustrating how input bands are processed by learnable Gabor kernels with varying orientations and frequencies.

Adaptive Feature Learning

By learning the phase $P$, the network adaptively blends low-frequency (cosine-like) and high-frequency (sine-like) information. This allows Gabor-Nets to automatically adapt to the local harmonic characteristics of HSI data. The architecture is composed of modular blocks, including Convolutional (CV) blocks and Fully Connected (FC) blocks, which allow it to be trained via standard backpropagation.

Structural details of the CV and FC blocks used in Gabor-Nets.

Experimental Results

The performance of Gabor-Nets was evaluated on three benchmark datasets: Pavia University (urban), Indian Pines (agricultural), and Houston (complex urban).

Key Findings

  • Superiority in Small Data Regimes: Gabor-Nets excel when training samples are scarce. On the Pavia University dataset with only 50 samples per class, Gabor-Nets achieved 95.91% accuracy, significantly outperforming standard CNNs at 85.52%.
  • Optimization Stability: By constraining the search space to physically meaningful features, Gabor-Nets avoid the "gradient vanishing" problem and reach high accuracy in fewer epochs than unconstrained models.
  • Scalability: Unlike standard CNNs, which often suffer from performance degradation as depth increases due to overfitting, Gabor-Nets remain stable and even improve with additional layers.

Conclusion

Gabor-Nets represent a significant step forward in HSI classification by combining the representational power of deep learning with the structural efficiency of Gabor filters. While the model occasionally produces over-smoothed results that can obscure tiny objects, its robustness to small training sets and its ability to produce clean, noise-resistant classification maps make it a powerful tool for remote sensing. Future work will focus on incorporating edge-preservation strategies to maintain sharp boundaries between land-cover classes.

Summary updated

Abstract

Recently, many convolutional neural network (CNN) methods have been designed for hyperspectral image (HSI) classification since CNNs are able to produce good representations of data, which greatly benefits from a huge number of parameters. However, solving such a high-dimensional optimization problem often requires a large amount of training samples in order to avoid overfitting. Additionally, it is a typical non-convex problem affected by many local minima and flat regions. To address these problems, in this paper, we introduce naive Gabor Networks or Gabor-Nets which, for the first time in the literature, design and learn CNN kernels strictly in the form of Gabor filters, aiming to reduce the number of involved parameters and constrain the solution space, and hence improve the performances of CNNs. Specifically, we develop an innovative phase-induced Gabor kernel, which is trickily designed to perform the Gabor feature learning via a linear combination of local low-frequency and high-frequency components of data controlled by the kernel phase. With the phase-induced Gabor kernel, the proposed Gabor-Nets gains the ability to automatically adapt to the local harmonic characteristics of the HSI data and thus yields more representative harmonic features. Also, this kernel can fulfill the traditional complex-valued Gabor filtering in a real-valued manner, hence making Gabor-Nets easily perform in a usual CNN thread. We evaluated our newly developed Gabor-Nets on three well-known HSIs, suggesting that our proposed Gabor-Nets can significantly improve the performance of CNNs, particularly with a small training set.
BibTeX
@misc{liu2019naive,
  author = {Chenying Liu and Jun Li and Lin He and Antonio Plaza and Shutao Li and Bo Li},
  title = {{Naive Gabor Networks for Hyperspectral Image Classification}},
  year = {2019},
  eprint = {1912.03991},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.03991},
}
Figure