作 者：
王炜超;杨玮;崔玉露;周鹏;王懂;李民赞
单 位：
中国农业大学现代精细农业系统集成研究教育部重点实验室
关键词：
土壤全氮;高光谱;CatBoost;图谱特征融合
摘 要：
针对高光谱技术应用于土壤养分定量检测中忽略彩色图像外部特征与土壤养分的内在关系的问题,结合土壤的光谱信息与图像特征构建一种图谱特征融合的土壤全氮含量预测模型,探究图谱特征融合对于土壤全氮含量的预测能力。通过实验室高光谱成像仪获取土壤样品的高光谱图像,从高光谱图像提取土壤的光谱信息与图像特征。使用无信息变量消除算法(Uniformative variable elimination, UVE)和竞争性自适应重加权采样算法(Competitive adaptive reweighted sampling, CARS)的联合算法对光谱信息进行特征波长的选择,选择后的特征波长作为土壤光谱信息;通过相关性分析选择与土壤全氮含量相关性较高的图像特征。将CatBoost(Categorical Boosting)算法应用到土壤全氮含量预测中,分别对基于单一光谱信息、单一图像特征和图谱特征融合对土壤全氮含量进行预测并比较。结果表明,UVE-CARS联合算法选取的特征波长为942、1 045、1 199、1 305、1 449、1 536、1 600 nm,与含氮基团的倍频吸收相吻合。与土壤全氮含量相关性较高的图像特征为角二阶矩、能量、惯性矩、灰度均值和熵。通过CatBoost算法建立的基于单一光谱信息特征波长的模型最终预测土壤全氮含量R~2为0.832 9,RMSE为0.203 3 g/kg;基于图像特征建立的模型最终预测土壤全氮含量R~2为0.801 7,RMSE为0.219 7 g/kg;基于图谱特征融合建立的模型最终预测土壤全氮含量R~2为0.866 8,RMSE为0.160 2 g/kg,预测精度均高于单一光谱特征和单一图像特征的预测精度,与基于单一光谱特征和单一图像特征相比,基于高光谱图谱特征融合的土壤全氮含量预测模型效果较好,为土壤全氮含量的预测研究提出一种新的方法。
译 名：
Prediction of Soil Total Nitrogen Based on CatBoost Algorithm and Fusion of Image Spectral Features
作 者：
WANG Weichao;YANG Wei;CUI Yulu;ZHOU Peng;WANG Dong;LI Minzan;Key Laboratory of Modern Precision Agriculture System Integration Research, Ministry of Education,China Agricultural University;
关键词：
soil total nitrogen;;hyperspectral;;CatBoost;;fusion of image spectral features
摘 要：
In order to solve the problem that the internal relationship between external features of color images and soil nutrients is ignored when hyperspectral technology is applied to quantitative detection of soil nutrients, a prediction model of soil total nitrogen content based on image and spectral features was constructed by combining the spectral information and image features of soil, and the prediction ability of image and spectral features fusion for soil total nitrogen content was explored. The hyperspectral images of soil samples were obtained by the laboratory hyperspectral imager, and the spectral information and image characteristics of soil were extracted from the hyperspectral images. The characteristic wavelength of spectral information was selected by using a joint algorithm of uniformative variable elimination(UVE) and competitive adaptive reweighted sampling(CARS), and the selected characteristic wavelength was used as soil spectral information. Through correlation analysis, image features with high correlation with soil total nitrogen content were selected. Categorical Boosting(CatBoost) algorithm was applied to the prediction of soil total nitrogen content, and the prediction of soil total nitrogen content based on single spectral information, single image feature and map feature fusion was compared. The results showed that the characteristic wavelengths selected by UVE-CARS joint algorithm were 942 nm, 1 045 nm, 1 199 nm, 1 305 nm, 1 449 nm, 1 536 nm and 1 600 nm, which were consistent with the frequency doubling absorption of nitrogen-containing groups. The image features with high correlation with soil total nitrogen content were angle second moment, energy, inertia moment, gray mean and entropy. The model based on the characteristic wavelength of single spectral information established by CatBoost algorithm finally predicted that the total nitrogen content of soil R~2 was 0.832 9 and RMSE was 0.203 3 g/kg, the model based on image features finally predicted that the total nitrogen content of soil R~2 was 0.801 7 and RMSE was 0.219 7 g/kg. And the model based on fusion of image and spectral features finally predicted that the total nitrogen content of the soil R~2 was 0.866 8, and RMSE was 0.160 2 g/kg, the prediction accuracy was higher than that of single spectral feature and single image feature. Compared with the prediction model based on single spectral feature and single image feature, the prediction model based on hyperspectral atlas feature fusion had better effect, which can provide a method for the prediction of soil total nitrogen content.
