联合注意力与混合卷积的高光谱地物识别研究

Research on hyperspectral ground object recognition based oncombined attention and mixed convolution

针对高光谱地物识别模型在提取空谱联合特征时,缺乏对空间特征有效关注的问题,提出了一种基于注意力机制和混合卷积神经网络的高光谱地物识别方法.该方法用三维CNN(3DCNN)以及二维CNN(2DCNN)对高光谱图像的空谱联合特征进行提取,并在二维卷积阶段引入了注意力机制,构建AFCNet地物识别模型,使得其在提取空谱联合特征的同时,实现对空间特征的有效关注和激活.所提模型使用带批归一化层(batch normalization,BN)的3D卷积核和2D卷积核,加快了模型的收敛速度,防止了过拟合现象的发生.相对于传统的卷积网络模型,所提模型提高了噪声抑制能力,得到了较好的地物识别效果,在Salinas和Pavia University&Center数据集上,取得了99.96%和99.87%的地物识别精度,验证了所提方法的有效性.

A novel method for high-precision ground object recognition is proposed in this study.The methodaddresses the issue of insufficient attention to spatial features when extracting joint spectral features inhyperspectral ground object recognition models.The proposed method employs a hybrid convolutional neuralnetwork architecture that combines the use of three-dimensional convolutional neural network(3DCNN)and two-dimensional convolutional neural network(2DCNN).Additionally,an attention mechanism is introduced in the 2Dconvolution stage to construct the AFCNet ground object recognition model.This enables effective attention andactivation of spatial features while extracting joint spectral features.To prevent overfitting and speed up theconvergence rate of the model,the proposed model employs batch normalization layers(BN)in conjunction with3D and 2D convolution kernels.The proposed method improves noise suppression capability and achievesexcellent ground object recognition results compared to traditional convolutional network models.To validate theproposed method,experiments were conducted on two publicly available hyperspectral datasets:Salinas and PaviaUniversity&Center datasets.The results showed that the proposed method achieved remarkable ground objectrecognition accuracies of 99.96%and 99.87%,respectively,demonstrating the effectiveness of the proposedmethod.