Improving Classification Accuracy of Hyperspectral Image Using Convolutional Neural Networks
|
Mahsa Tekyeh-Nejad , Ata Allah Ebrahimzadeh * , Maliheh Ahmadi |
Babol Noshirvani University of Technology |
|
Abstract: (1563 Views) |
Hyperspectral image classification is a crucial aspect of remote sensing image analysis. Deep learning methods have been successfully used to classify remote sensing data. In recent years, convolutional neural networks (CNNs) have been significantly used in hyperspectral image classification, which has tried to overcome the computational and processing challenges of hyperspectral data. By increasing the number of parameters and layers of convolutional neural networks, their efficiency in solving complex problems decreases. For this reason, in this article, a new architecture of convolutional neural networks has been introduced, this network has a good performance and reduces the computing time.
In this paper, we introduce a novel CNN that utilizes spectral-spatial information as input and employs principal component analysis (PCA) to reduce spectral bands. To prevent overfitting, we combine batch normalization and dropout techniques. Our two-dimensional CNN includes convolutional layers, pooling layers, and fully connected layers. We also incorporate PCA and patch selection to enhance the accuracy of our model.
To evaluate the effectiveness of our proposed model, we conducted experiments on three datasets: Indian Pines, Pavia University, and Salinas. Our simulation results demonstrate that our model achieves a classification accuracy of 100%, with less training time and complexity than existing models.
|
|
Keywords: Hyperspectral image classification, Convolutional neural network, Principal component analysis, Choose the right patches |
|
Full-Text [PDF 1636 kb]
(384 Downloads)
|
Type of Study: Research |
Subject:
RS Received: 2022/10/15 | Accepted: 2023/06/3 | ePublished ahead of print: 2023/06/21 | Published: 2023/07/9
|
|
|
|
|
Send email to the article author |
|