基于卷积神经网络和稳定性选择的农作物植被光谱分析方法

SPECTRAL ANALYSIS OF CROP VEGETATION BASEDCONVOLUTIONAL NEURAL NETWORK AND STABILITY SELECTION

光谱特征提取是遥感植被精细化识别研究的关键.现有光谱分析方法大多无法直接高效的提取特征波段用于分类,往往依赖特征选择和预处理操作,目标识别精度不高.本文整合卷积神经网络(CNN)和稳定性选择改进光谱分析方法实现统一高效的光谱分析,以玉米、大豆、豆角、葡萄、大枣、辣椒、秋葵、芥蓝、韭菜和草皮10种农作物植被叶片光谱为实验对象,构建适用于农作物分类的CNN模型,获得植被分类结果.利用稳定性选择方法可视化CNN卷积-池化过程的特征选择结果,获得表征不同植被生化参数的特征波段.结果表明:1)改进的光谱分析方法很好的适用于光谱识别,分类准确率维持在97%~100%之间;2)该方法对光谱预处理依赖性最小,对光谱识别表现出较强的鲁棒性和泛化能力;3)特征波段的可视化结果证明了改进的光谱分析方法能够精准的提取农作物植被的敏感特征波段,间接说明了CNN模型卷积池化操作能够准确地提取光谱重要特征波段用于农作物光谱分类.

Spectral feature extraction is the key to the study of the plant accurate identification,which use remote sensing spectrum.Most existing spectral analysis methods cannot directly extract effective feature bands for classification.These models often rely on feature selection and preprocessing,and cannot obtain high classification accuracy.For this reason,the convolution neural network(CNN)and stability feature selection are integrated to improve spectral analysis method for achieving unified and efficient spectral analysis.We selected corn,soybean,long bean,grape,jujube,chilli,okra,Chinese kale,leek and turf 10 crop plants for field spectral measurement(to build the model of CNN)and took these spectra as experimental objects.In order to find the best parameter values,the Random Grid Search Cross-Validation framework(RGS-CV)was used to obtain the best CNN model for crops classification.Stability selection method is used to train the output layer of CNN model.The output of convolution-pooled layer was used as the input of stability selection method to calculate the potential spectral feature band.Finally,crop vegetation recognition results were compared with partial least squares(PLS-DA)and softmax.The experimental results showed that:1)the CNN is suitable for the hyperspectral identification of the crops,where classification accuracy remains between 97%and 100%,and the classification accuracy was higher than PLS-DA and softmax under the original spectrum and different preprocessing conditions;2)the CNN model has the least dependence on spectral preprocessing and shows strong robustness and generalization ability for crops hyperspectral recognition;3)the visualization results of characteristic bands proved that the improved spectral analysis method can accurately extract the sensitive characteristic bands of crop vegetation.Indirectly,it showed that CNN convolution pooling can accurately extract the hyperspectral important feature bands for crop classification and improve the classification accuracy.