A Multi-branch Feature Fusion Model Based on Convolutional Neural Network for Hyperspectral Remote Sensing Image Classification

Hyperspectral image classification constitutes a pivotal research domain in the realm of remote sensing image processing. In the past few years, convolutional neural networks (CNNs) with advanced feature extraction capabilities have demonstrated remarkable performance in hyperspectral image classification. However, the challenges faced by classification methods are compounded by the difficulties of "dimensional disaster" and limited sample distinctiveness in hyperspectral images. Despite existing efforts to extract spectral spatial information, low classification accuracy remains a persistent issue. Therefore, this paper proposes a multi-branch feature fusion model classification method based on convolutional neural networks to fully extract more effective and adequate high-level semantic features. The proposed classification model first undergoes PCA dimensionality reduction, followed by a multi-branch network composed of three-dimensional and two-dimensional convolutions. Convolutional kernels of varying scales are utilized for multi-feature extraction. Among them, the 3D convolution not only adapts to the cube of hyperspectral data but also fully exploits the spectral-spatial information, while the 2D convolution learns deeper spatial information. The experimental results of the proposed model on three datasets demonstrate its superior performance over traditional classification models, enabling it to accomplish the task of hyperspectral image classification more effectively.