利用多种回归模型对比估算琯溪蜜柚叶片钾素含量

Comparative Estimation of Potassium Contents in Guanxi Honey Pomelo Leaves by Multiple Regression Models

钾素是蜜柚营养三要素之一,是准确诊断和定量评价生长状况的重要指标,建立合适的蜜柚叶片钾素含量高光谱估算模型,为实现快速、无损、精确的钾素含量估测提供依据。基于蜜柚叶片高光谱数据和钾素含量实测数据,首先分析蜜柚叶片钾素含量与原始及一阶微分光谱的相关性,然后分析与敏感波段植被参数的相关性,并找出相关性较好的光谱参数,建立蜜柚叶片钾素含量偏最小二乘回归模型(PLS)、BP神经网络回归模型(BPNN)、随机森林回归模型(RF)和支持向量机回归模型(SVM),并确定蜜柚叶片钾素含量最佳估算模型。在513~598 nm和699~735 nm处,蜜柚叶片钾素含量与原始光谱反射率呈显著负相关,最大负相关系数分别为‒0.47(554 nm)和‒0.45(715 nm)。在507~552 nm和691~711 nm处,蜜柚叶片钾素含量与一阶微分光谱反射率呈显著负相关,最大负相关系数分别为‒0.54(528 nm)和‒0.53(702 nm);在557~655 nm处,二者呈显著正相关,最大正相关系数为0.58(579 nm)。选择554、715、528、579、702 nm构建光谱参数,建立差值植被指数(DVI_(λ1,λ2))、比值植被指数(RVIλ1,λ2)和归一化植被指数(NDVI_(λ1,λ2))等,其中与蜜柚叶片钾素含量相关性较好的光谱参量为NDVI′_(579,702)、RVI_(554,715)、RVI′_(528,579)、R′_(579)。建立PLS、BPNN、RF和SVM等4种回归模型估算蜜柚叶片钾素含量并进行对比验证,4种估算模型的决定系数(R2)分别为0.72、0.74、0.84和0.81,均方根误差(RMSE)分别为2.44、2.28、1.49和1.61;相对误差(RE)分别为9.95%、9.01%、7.84%和8.01%。验证模型的R2分别为0.79、0.84、0.85和0.82,RMSE分别为1.69、1.48、1.34和1.41,RE分别为8.47%、7.70%、6.12%和6.35%,RF估算模型精度稍高于PLS、BPNN和SVM估算模型。

Potassium(K)is one of the three nutrient elements of honey pomelo,which is an important index for accurate diagnosis and quantitative evaluation of growth status.The hyperspectral estimation model of K contents in honey pomelo leaves was established to provide basis for rapid,non-destructive and accurate estimation of K contents.Based on the hyperspectral data of pomelo leaves and the measured data of K contents,this study first analyzed the correlation between the K contents of pomelo leaves and the original and first-order differential spectra,then analyzed the correla-tion between sensitive band vegetation index and the K contents of pomelo leaves,and found out the spectral parameters with good correlation with the K contents of pomelo leaves,then the partial least squares regression model(PLS),BP neural network regression model(BPNN),random forest regression model(RF)and support vector machine regression model(SVM)of pomelo leaves K contents were established,and the best estimation model of K contents in pomelo leaves was determined.In the measured band of 350‒1050 nm,the spectral reflectance of pomelo leaves decreased with the increase of K contents.In 513‒598 nm and 699‒735 nm,it reached a significant negative correlation level,the maximum negative correlation coefficient was‒0.47(554 nm)and‒0.45(715 nm),respectively.In 507‒552 nm and 691‒711 nm,potassium in pomelo leaves reached a significant negative correlation level with the first-order spectral reflectance,the maximum negative correlation coefficient was‒0.54(528 nm)and‒0.53(702 nm).In 557‒655 nm,it reached a significant positive correlation level,and the maximum positive correlation coefficient was 0.58(579 nm).554,715,528,579,702 nm were selected to construct the spectral parameters and establish the difference vegetation index(DVI_(λ1,λ2)),ratio vegetation index(RVIλ1,λ2)and normalized difference vegetation index(NDVI)λ1,λ2).NDVI¢579,702,RVI_(554,715),RVI¢528,579,R¢579 were the spectral parameters that had good correlations with the K contents of pomelo leaves.Four regression models such as PLS,BPNN,RF and SVM were established to estimate K content in pomelo leaves and verified.R2,RMSE and RE of the estimation model of pomelo leaves K contents established by RF method was 0.84,1.49 and 7.84%,respectively.R2,RMSE and RE of the estimation model by SVM method were 0.81,1.61 and 8.01%respectively.While R2,RMSE and RE of the estimation model by BPNN method were 0.74,2.28 and 9.01%respectively;R2,RMSE and RE of the estimation model by PLS method were 0.72,2.44 and 9.95%respectively.R2 of the validation model of PLS,BPNN,RF and SVM methods were 0.79,0.84,0.85 and 0.82 respectively.Compared with PLS,BPNN and SVM,RF had higher R2,lower RMSE and lower RE,indicating that the accuracy of RF based K contents estimation model was higher than that of PLS,BPNN and SVM.Through the comparison of four hyperspectral estimation models for K contents in Guanxi honey pomelo leaves,the accuracy of random forest estimation model was higher than that of PLS,BPNN and SVM.