基于高光谱数据的高寒草甸主要毒草分类技术研究

Technology for identifying poisonous plants in alpine meadow based on hyperspectral data

摘要:

毒草种类识别是开展毒草危害面积调查、危害程度评估以及科学防治的前提。研发快速、高效和适用于大范围的毒草种类识别技术对于退化草地生态修复具有重要意义。本研究利用SOC710VP近红外高光谱成像仪,获取了高寒草甸11种主要毒草的光谱数据。采用Savitzky-Golay平滑处理、导数变换、归一化变换、均值中心化和对数变换等方式对高光谱数据进行预处理,利用主成分分析(PCA)方法对预处理数据进行降维处理后,分别用随机森林(RF)、支持向量机-径向核函数(SVM-RBF)、K临近算法(Knn)、朴素贝叶斯(NB)、决策树(DT)等算法对降维后光谱数据进行分类预测,以混淆矩阵总体精度为检验标准,筛选可用于高寒草甸主要毒草的近红外高光谱数据识别算法。结果表明: 1)6种数学变换处理中,对数变换后样本光谱反射值差异最大;2)对数变换、平滑处理、均值中心化变换经PCA降维后PC1和PC2累计方差贡献率＞85%;3)5种分类算法中,SVM-RBF算法分类精度最高,分类精度达到99.35%;4)使用前8个主成分分类时,分类精度SVM-RBF＞RF＞NB＞Knn＞DT。

Abstract:

The identification of poisonous plants is the prerequisite for investigating their distribution,assessing the damage degree as well as their scientific control.Developing a fast and efficient technology for identifying poisonous plants in large scales is of great significance for the restoration of degrading grassland.In this study,SOC710VP near-infrared hyperspectral imager was used to obtain the spectral data of 11 major poisonous plants in alpine meadows.The hyperspectral data were preprocessed using Savitzky-Golay smoothing processing,derivative transformation,normalization transformation,mean centralization and logarithmic transformation.Following that,principal component analysis(PCA) was used to reduce the dimension of the preprocessed data. The classification models including random forest(RF),support vector machine-radial kernel function(SVM-RBF),K-nearest algorithm(Knn),naive Bayes(NB) and decision tree(DT),were then used for classification test.The overall accuracy of the confusion matrix was employed as the test standard to select a suitable classification model for identifying the primary poisonous plants in alpine meadow.The results showed that among six mathematical data transformation,the logarithmic transformation was the best showing the largest difference in spectral reflection among samples.The cumulative variance contribution rate of PC1 and PC2 in PCA by logarithmic transformation,smoothing and mean centralization transformation were more than 85%.Among the five classification algorithms,the SVM-RBF algorithm had the highest classification accuracy of 99.35%.When the first eight principal components were used for classification,the rank of classification accuracy followed the order of SVM-RBF＞RF＞NB＞Knn＞DT.