基于多尺度级联卷积神经网络的高光谱图像分析

Analysis of hyperspectral image based on multi-scale cascaded convolutional neural network

在样本有限情况下对象级别的叶片高光谱图像建模中,提出了多尺度三维和一维级联卷积神经网络模型。首先,在三维卷积神经网络(3D-CNN)中嵌入扩张卷积增大卷积核感受野,构建了多尺度3D-CNN,提取和融合不同尺度的光谱-空间联合特征,在不增加网络参数的情况下提升了模型性能。然后,对最优多尺度3D-CNN网络级联一维卷积神经网络(1D-CNN),进一步降低计算复杂度和过拟合程度。最后,在罗勒叶片叶绿素含量回归和辣椒叶片干旱胁迫识别两类数据集上进行最优网络框架探究并对比了一系列基准CNN模型。结果表明,对于叶片高光谱图像回归和分类,本文模型均能在小样本条件下有效提升模型泛化性能并降低计算复杂度。

In object-level hyperspectral image modeling with limited leaf samples,a multi-scale three dimensional-one dimensional cascaded convolution neural network model was proposed.Firstly,in 3D convolution neural network(3D-CNN),the dilated convolution was embedded to increase the receptive field of convolution kernel,and a multi-scale 3D-CNN network was constructed to extract and fuse spectral-spatial joint features of different scales,to improve the model performance without increasing the number of network parameters.Then,the one-dimensional convolutional neural network(1D-CNN)was cascaded to the optimal multi-scale 3D-CNN network,to further reduce the computational complexity and overfitting degree of model.Finally,on two datasets of chlorophyll content regression for basil leaf and drought stress recognition for pepper leaf,the optimal network architecture was explored,with comparison to a series of baseline CNN models.Experimental results showed that the proposed model can effectively improve the generalization performance and reduce the computational complexity for both regression and classification tasks of leaf hyperspectral images under the condition of small samples.