基于BiPLS-SPA優選近紅外光譜的木材基本密度預測1)——以柞木為例

張怡卓 涂文俊 李超 潘屾

(東北林業大學，哈爾濱，150040)

?

基于BiPLS-SPA優選近紅外光譜的木材基本密度預測1)
——以柞木為例

張怡卓 涂文俊 李超 潘屾

(東北林業大學，哈爾濱，150040)

以柞木為研究對象，以900～1 700 nm的近紅外光譜儀獲取木材表面近紅外光譜數據，對89個柞木樣本進行檢測，其中58個組成校正集，31個為預測集。首先，采集樣本徑切面光譜數據，并利用SG平滑對光譜數據進行預處理；然后，利用反向區間偏最小二乘(BiPLS)選出均方根誤差最小的波長區間組合；再利用連續投影算法(SPA)進一步選擇出波長特征；最后，以優選出的波長特征作為輸入，建立偏最小二乘法回歸模型，確定出木材基本密度與近紅外光譜之間的聯系。BiPLS算法將光譜劃分區間劃分為10時，均方根誤差最小，其最佳區間組合為[3 5 6 7 9]，變量個數由全光譜117個降至59個；應用SPA算法二次降維，變量個數降至6個，降低變量信息的冗余，減少了變量個數，提高了建模的速度和效率。BiPLS-SPA模型較PLS、iPLS、BiPLS、SPA-PLS具有更高的相關系數，更小的均方根誤差，柞木基本密度預測相關系數為0.925，預測均方根誤差為0.010 4，相對分析誤差為2.83。

木材；柞木；基本密度；近紅外；偏最小二乘法；連續投影算法

WithXylosmaracemosumas the research object, 900-1 700 nm near-infrared spectrometer was used to obtain wood surface spectral data. The 89X.racemosumsamples were detected, of which 58 composed the calibration set, and 31 were used for the prediction set. Firstly, the diameter section spectral data was acquired and preprocessed by SG smoothing method; Secondly, backward interval partial least squares (BiPLS) was implemented to divide the spectrum into several wavelength interval, and intervals with the smallest RMSE were selected as a data combination; thirdly, successive projections algorithm (SPA) was chosen to select the wavelength characteristics from the data combination; Then, using optimized characteristics as the input variable, partial least squares regression model can be established and finally the correlation between the near infrared spectrum and wood basic density was built. The RMSECV had minimum value when the spectrum was divided into 10 intervals, and the optimum interval combination was [3 5 6 7 9], and the number of variables dropped from 117 to 6. Consequently, the number of variables were reduced and the modeling speed was increased. BiPLS-SPA model has a higher correlation coefficient than the PLS, iPLS, BiPLS, SPA-PLS method. The prediction correlation coefficient ofX.racemosumbasic density is 0.925, with the RMSEP of 0.010 4, and the RPD of 2.83.

近紅外光譜分析具有無損、安全的特點，國內外學者已利用近紅外光譜分析技術開展了木材性質的研究[1-5]。由于近紅外光譜區域存在與木材基本密度不相關或者相關性較小的特征，在建模過程中一定程度的引入了冗余信息，導致增加了偏最小二乘回歸過程的預測方差，降低了模型精度。……