基于高光谱特征的松材线虫岭回归估测模型研究

Ridge Regression Model for Estimating Pine Wilt Disease Based on Hyperspectral Characteristics

以2017年6—8月获取的重庆永胜林场马尾松光谱反射率为数据源,对绿光区(490～560 nm)、黄光区(560～590 nm)、红光区(620～680 nm)、红边(680～780 nm)、近红外区(780～1 100 nm)最大反射率和反射率总和、绿峰(500～670 nm)反射高度、红谷(560～760 nm)吸收深度等14个高光谱特征参数进行岭迹分析,筛选出非共线性特征参数,构建松材线虫岭回归估测模型。结果表明:红边和近红外区反射率最大值、红边和近红外区反射率总和、红谷吸收深度岭迹曲线变化稳定且不趋于零,可用于岭回归建模。当岭迹参数k=0. 2时,上述5个高光谱特征参数岭迹趋于稳定,根据k值计算岭回归系数,构建松材线虫岭回归估测模型。模型决定系数R2为0. 868 6,均方根误差RMSE为0. 273 5,平均估测精度为87. 15%,可为松材线虫病害早期监测和防治研究提供技术支持。

Pine wilt disease (PWD) caused by the pine wood nematode, Bursaphelenchus xylophilus , is considered as the most destructive forest-invasive alien species and may cause serious economic losses. A ridge regression model was proposed based on the hyperspectral characteristics to estimate the degrees of pine wilt disease for Pinus massoniana in Yongsheng forest of Chongqing, Southwest China. The spectral reflectance and quantitated pet levels for Pinus massoniana were measured from June to August 2017. And then the ridge trace analysis was operated on 14 spectral characteristics, which covered maximum and sum of reflectance ranging in green region (490~560 nm), yellow region (560~590 nm), red region (620~680 nm), red edge (680~780 nm), near-infrared region (780~1 100 nm), as well as the reflectance height of green peak (500~670 nm) and absorption depth of red valley (560~760 nm). Furthermore, the hyperspectral characteristic parameters with less collinearity were selected to construct the estimation model of PWD with ridge regression. The results demonstrated that ridge trace curves for the maximum of reflectance in red edge, near-infrared region, the sum of reflectance in the red edge, near-infrared region, as well as absorption depth of red valley were stable, which were not close to zero. Therefore, those five spectral characteristics could be considered in ridge regression modeling;when the ridge trace parameter k was 0.2, the ridge traces of the above five hyperspectral characteristic parameters became stable, and then the ridge regression coefficients were calculated. Finally, a regression estimation model of PWD was built with determination coefficient R^2 of 0.868 6, root-mean-square error (RMSE) of 0.273 5, and average estimation accuracy of 87.15%. The research provided both scientific support and application reference for monitoring forest pet disease with remote sensing technology.