基于机器学习的油茶叶片钾含量估算模型构建  

Construction of potassium content estimation model for Camellia oleifera leaves based on machine learning

作　　者：唐雪海[1] 燕李鹏 傅根深 匡帆 窦敏 黄庆丰[1] 欧强新 TANG Xue-hai;YAN Li-peng;FU Gen-shen;KUANG Fan;DOU Min;HUANG Qing-feng;OU Qiang-xin(School of Forestry and Landscape Architecture of Anhui Agricultural University,Anhui Provincial Key Laboratory of Forest Resources and Silviculture,Hefei 230036,China)

机构地区：[1]安徽农业大学林学与园林学院,安徽省林木资源培育重点实验室,安徽合肥230036

出　　处：《中国油料作物学报》2025年第2期488-501,共14页Chinese Journal of Oil Crop Sciences

基　　金：国家自然科学基金(32171783)。

摘　　要：为实现大面积油茶林生长遥感监测,构建适用于长林系列油茶钾含量估算模型,利用高光谱进行叶片钾(LKC)无损监测,探明油茶叶片钾含量与冠层光谱的响应关系。使用多元散射校正和Savitzky-Golay卷积一阶求导(SGFD)对长林系列油茶的冠层光谱进行预处理,建立多波段光谱指数组合,构建LKC最优估算模型。结果表明:LKC与原始光谱的响应在绿光和红光波段敏感区间的光谱反射率与叶片钾含量为负相关,反映养分含量变化对光合色素的整体影响;预处理效果上,SGFD整体优于多元散射校正,预处理与光谱指数的组合效果会随光谱维度的增加发生变化,冠层尺度下LKC与光谱特征的绝对值最大相关系数为0.62;混合变量选择策略VCPA-IRIV(变量组合集群分析VCPA和迭代保留信息变量IRIV的组合)对光谱变换特征具有99%以上的变量空间压缩率,有效提升了估算模型精度,经多元散射校正和SGFD预处理后的保留变量数增加,其中两波段和三波段光谱指数在入选波长组合位置上具有强弱光谱信号结合的特点。最适LKC模型是SGFD-NDSI-BPNN,R_(P)^(2)=0.84,RMSEP=0.35 g/kg,RPD=2.56。本文构建的长林系列油茶林LKC估算模型,可为大面积油茶林生长的遥感监测提供依据。To realize remote sensing monitoring of large-scale Camellia oleifera forest growth and establish a leaf potassium content(LKC)estimation model for Changlin series C.oleifera,non-destructive monitoring of LKC was conducted by hyperspectral method to investigate the response relationship between LKC and canopy spectrum.This paper took Changlin series C.oleifera as research materials,used multiplicative scatter correction and SGFD(Savitzky-Golay first derivative)to preprocess canopy spectrum,and constructed multi-band spectral indices to establish optimal LKC estimation model.Results showed that spectral reflectance in sensitive regions of green-and red-wavelengths were negatively correlated with LKC,reflecting the overall effect of nutrient content changes on photosynthetic pigments.The processing effect of SGFD was better than that of multiplicative scatter correction.The combined effect of pretreatment and spectral indices was different with the increase of spectral dimension.The maximum absolute correlation coefficient between LKC and spectral features was 0.62 at canopy scale.VCPA(variable combination population analysis)-IRIV(iteratively retaining informative variables)had a variable space compression ratio of more than 99% for spectral transformation features,which effectively improved the accuracy of the estimation models.The number of retained variables increased after MSC and SGFD pretreatment.In terms of selected wavelength combination positions,the two-band and three-band spectral indices had the characteristics of combining strong and weak spectral signals.The optimal LKC model was SGFD-NDSI-BPNN,R_(P)^(2)=0.84,RMSEP=0.35 g/kg,RPD=2.56.This study expanded the application of hyperspectral technology and clarified the relationship between LKC and canopy spectrum.The LKC estimation models of Changlin series C.oleifera were constructed,which provides a theoretical basis for remote sensing monitoring of C.oleifera forest growth in large areas.

关 键 词：油茶 叶片钾含量 光谱指数 多元散射校正 Savitzky-Golay卷积一阶求导 养分监测 

分 类 号：S565.9[农业科学—作物学] O657.3[理学—分析化学]