研究目的
To propose hybrid neural networks (HNN) for extracting both spatial and spectral features from hyperspectral images for classification, aiming to improve classification performance over traditional and other deep learning-based methods.
研究成果
The proposed hybrid convolutional neural networks achieve excellent performance in hyperspectral image classification by effectively combining spatial and spectral features. The method outperforms other state-of-the-art deep learning methods, especially when training samples are limited.
研究不足
The computational complexity of the proposed method is a limitation, and future work will focus on reducing this complexity in end-to-end networks.
1：Experimental Design and Method Selection:
The study employs a two-branch hybrid convolutional neural network to jointly extract spatial-spectral features in an end-to-end framework. It includes batch normalization, weighted convolution layers, and Gaussian pooling layers.
2：Sample Selection and Data Sources:
Two hyperspectral datasets, Indian Pines (IP) and Pavia Center (PC), are used for performance evaluation.
3：List of Experimental Equipment and Materials:
The datasets were collected by Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) and Reflective Optical System Imaging Spectrometer (ROSIS).
4：Experimental Procedures and Operational Workflow:
The proposed networks consist of spatial structure layers, spatial contextual layers, and spectral layers. The training strategy involves optimizing parameters through gradient descent and back-propagation algorithm.
5：Data Analysis Methods:
Performance is evaluated using overall accuracy (OA), average accuracy (AA), and kappa coefficient (κ).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
