[Paper Review] Hyperspectral Images Classification Based on Multi-scale Residual Network

Xiangdong Zhang, Wang Tengjun+1 more|arXiv (Cornell University)|Apr 26, 2020

Remote-Sensing Image Classification21 references4 citations

TL;DR

This paper proposes a multi-scale residual network (MSRN) for hyperspectral image classification that enhances feature extraction by integrating multi-scale convolutional branches within residual blocks, enabling effective fusion of spatial and spectral features. The method achieves 99.07% and 99.96% overall accuracy on the Indian Pines and Pavia University datasets, respectively, outperforming existing methods, especially under small training data settings through dynamic learning rate, batch normalization, and dropout for improved generalization and speed.

ABSTRACT

Because hyperspectral remote sensing images contain a lot of redundant information and the data structure is highly non-linear, leading to low classification accuracy of traditional machine learning methods. The latest research shows that hyperspectral image classification based on deep convolutional neural network has high accuracy. However, when a small amount of data is used for training, the classification accuracy of deep learning methods is greatly reduced. In order to solve the problem of low classification accuracy of existing algorithms on small samples of hyperspectral images, a multi-scale residual network is proposed. The multi-scale extraction and fusion of spatial and spectral features is realized by adding a branch structure into the residual block and using convolution kernels of different sizes in the branch. The spatial and spectral information contained in hyperspectral images are fully utilized to improve the classification accuracy. In addition, in order to improve the speed and prevent overfitting, the model uses dynamic learning rate, BN and Dropout strategies. The experimental results show that the overall classification accuracy of this method is 99.07% and 99.96% respectively in the data set of Indian Pines and Pavia University, which is better than other algorithms.

Motivation & Objective

• To address low classification accuracy in hyperspectral image analysis when using small training datasets.
• To improve feature representation by fusing multi-scale spatial and spectral information.
• To enhance model generalization and training efficiency through regularization techniques.
• To develop a deep learning architecture that outperforms existing methods on benchmark hyperspectral datasets.

Proposed method

• Introduces a multi-scale residual block that incorporates parallel convolutional branches with different kernel sizes to capture diverse receptive fields.
• Fuses features from multiple scales within each residual block to enhance spatial-spectral representation learning.
• Employs batch normalization (BN) and dropout layers to reduce overfitting and stabilize training.
• Uses dynamic learning rate scheduling to accelerate convergence and improve optimization.
• Applies global average pooling and fully connected layers for final classification.
• Trains the model end-to-end on hyperspectral image datasets using stochastic gradient descent.

Experimental results

Research questions

• RQ1 Can multi-scale convolutional branches within residual blocks improve feature extraction for hyperspectral image classification?
• RQ2 How does the proposed MSRN architecture perform compared to state-of-the-art methods on small-sample hyperspectral datasets?
• RQ3 To what extent do batch normalization and dropout improve generalization in low-data regimes?
• RQ4 Does the fusion of multi-scale spatial and spectral features lead to higher classification accuracy?
• RQ5 Can dynamic learning rate scheduling enhance training efficiency and model convergence?

Key findings

• The proposed MSRN achieves 99.07% overall classification accuracy on the Indian Pines dataset, surpassing previous methods.
• On the Pavia University dataset, the model attains 99.96% overall accuracy, demonstrating superior performance.
• The model shows robustness under small training data conditions due to effective regularization with batch normalization and dropout.
• The integration of multi-scale convolutional branches significantly improves feature representation and classification accuracy.
• Dynamic learning rate scheduling contributes to faster convergence and improved training stability.
• The ablation study confirms that multi-scale feature fusion and regularization techniques are key contributors to the model's high performance.