[Paper Review] 1D-Convolutional Capsule Network for Hyperspectral Image Classification

Hai‐Tao Zhang, Ling‐Guo Meng+6 more|arXiv (Cornell University)|Mar 23, 2019

Remote-Sensing Image Classification42 references19 citations

TL;DR

This paper proposes 1D-ConvCapsNet, a lightweight 1D-convolutional capsule network that separately extracts spatial and spectral features for hyperspectral image (HSI) classification. By using local connections and a capsule-wise constraint window, it reduces parameters and training cost while maintaining high accuracy, outperforming both 3D-CNNs and standard CapsNets in speed and efficiency with comparable or superior classification accuracy on three benchmark datasets.

ABSTRACT

Recently, convolutional neural networks (CNNs) have achieved excellent performances in many computer vision tasks. Specifically, for hyperspectral images (HSIs) classification, CNNs often require very complex structure due to the high dimension of HSIs. The complex structure of CNNs results in prohibitive training efforts. Moreover, the common situation in HSIs classification task is the lack of labeled samples, which results in accuracy deterioration of CNNs. In this work, we develop an easy-to-implement capsule network to alleviate the aforementioned problems, i.e., 1D-convolution capsule network (1D-ConvCapsNet). Firstly, 1D-ConvCapsNet separately extracts spatial and spectral information on spatial and spectral domains, which is more lightweight than 3D-convolution due to fewer parameters. Secondly, 1D-ConvCapsNet utilizes the capsule-wise constraint window method to reduce parameter amount and computational complexity of conventional capsule network. Finally, 1D-ConvCapsNet obtains accurate predictions with respect to input samples via dynamic routing. The effectiveness of the 1D-ConvCapsNet is verified by three representative HSI datasets. Experimental results demonstrate that 1D-ConvCapsNet is superior to state-of-the-art methods in both the accuracy and training effort.

Motivation & Objective

• To address the high computational cost and parameter complexity of 3D-CNNs in hyperspectral image (HSI) classification due to the high dimensionality of HSI data.
• To overcome the limitations of standard capsule networks (CapsNets), such as high memory usage and slow training, in HSI applications.
• To develop a lightweight, efficient deep learning model that maintains high classification accuracy even with limited labeled training samples.
• To enable faster convergence and reduced training effort while preserving spatial-spectral feature representation in HSI data.

Proposed method

• The model uses separate 1D convolutional layers to extract spatial and spectral features independently, reducing parameter count compared to 3D-convolutional networks.
• A primary capsule layer with local connections and weight sharing is applied to reduce the number of parameters and computational load.
• A capsule-wise constraint window mechanism is introduced to limit the receptive field of capsules, further minimizing parameter redundancy.
• Dynamic routing is used to propagate activity vectors between capsule layers, enabling hierarchical representation learning of part-whole relationships.
• Principal component analysis (PCA) whitening is applied to input data to improve feature representation and reduce redundancy.
• The network is trained end-to-end using backpropagation with a margin loss function to optimize classification performance.

Experimental results

Research questions

• RQ1 Can a 1D-convolutional capsule network achieve competitive classification accuracy on hyperspectral images with significantly reduced training time and parameter count compared to 3D-CNNs?
• RQ2 How does the proposed 1D-ConvCapsNet perform under low-data regimes, such as 5% or 10% labeled training samples?
• RQ3 To what extent does the capsule-wise constraint window reduce parameter count and computational complexity in capsule networks for HSI classification?
• RQ4 How does the separation of spatial and spectral feature extraction compare to joint 3D-convolution in terms of accuracy and efficiency?

Key findings

• 1D-ConvCapsNet achieved 97.89% overall accuracy on the Indian Pines dataset with 10% training samples, and 94.64% with only 5% training samples, demonstrating strong generalization under data scarcity.
• On the University of Pavia dataset, the model reached 99.28% accuracy with 10% training samples and 98.47% with 5%, showing robustness to limited data.
• On the Salinas dataset, 1D-ConvCapsNet achieved 99.56% accuracy with 10% training samples and 99.22% with 5%, indicating excellent performance on small training sets.
• The model reduced training time by 93-96% and parameter count by 93-96% compared to standard CapsNet across all three datasets, with training times of 402s (Indian Pines), 432s (University of Pavia), and 2,146s (Salinas).
• 1D-ConvCapsNet converged faster than c-CNN and achieved higher validation accuracy, with convergence speed comparable to CapsNet but with significantly lower computational cost.
• Classification maps from 1D-ConvCapsNet showed fewer misclassifications and noise compared to c-CNN, with results closer to ground truth, especially in complex or noisy regions like Corn-no and Soy-no in the Indian Pines dataset.