Abstract:

【Objective】 To explore the accuracy of soil organic matter and total nitrogen estimation model with different modeling methods and to establish a fast and stable estimation model, so as to provide scientific basis for precision fertilization of modern agricultural production. 【Methods】 Taking the cultivated soil from Bortala Mongolia Autonomous Prefecture as the research object, the ASD Field4 ground object spectrometer was used to measure the spectrum of the treated soil samples in the dark room. The original spectrum was processed by breakpoint fitting and Savitzky-Golay (S-G) smoothing filtering correction. First derivative (FD), first derivative of logarithm ((lgR)'), first derivative of reciprocal ((1/R)') and multipication scatter correction (MSC) were performed on the corrected spectrum (R). The correlation analysis of the above five forms of spectra with soil organic matter and total nitrogen content was carried out to screen the characteristic bands. Based on the characteristic bands, partial least squares regression (PLSR), BP neural network (BP) and random forest (RF) were used to establish the estimation models of soil organic matter and total nitrogen, and the accuracy and stability of the models were evaluated. 【Results】 After different transformations, the correlation coefficients between the spectra and soil organic matter and total nitrogen increased, and the characteristic bands were more obvious. The first derivative transformation of the first derivative and the reciprocal was better than those of other transformations. The FD-PLSR model had the highest accuracy in predicting organic matter, with Rv² and RPD of 0.89 and 2.63, respectively. The (1/R)'-PLSR model had the highest accuracy in predicting soil total nitrogen, with Rv² and RPD of 0.83 and 2.42, respectively. 【Conclusion】 Based on hyperspectral technology and machine learning model, the estimation of soil organic matter and total nitrogen in cultivated land of Bozhou can be realized.

Key words: hyperspectral; soil organic matter; soil total nitrogen; spectral estimation; partial least squares regression; random forest

摘要：

【目的】 研究土壤高光谱数据经不同形式变换后与不同建模方法构建土壤有机质与全氮估测模型的精度,建立快速、稳定的估测模型,为现代化农业生产的精准施肥提供科学依据。【方法】 以新疆博尔塔拉蒙古自治州(简称博州)耕地土壤为研究对象,在暗室中使用ASD Field4地物光谱仪测量处理后的土壤样品光谱。将原始光谱进行断点拟合与Savitzky-Golay(S-G)平滑滤波校正处理,对校正后光谱(R)进行一阶导数(First Derivative,FD)、对数的一阶导数(First derivative of logarithmic,(lgR)’)、倒数的一阶导数(First derivative of reciprocal,(1/R)’)、多元散射校正(Multipication scatter correction,MSC)4种变换,分析5种光谱数据与土壤有机质和全氮含量,筛选特征波段,基于特征波段运用偏最小二乘回归(PLSR)、BP神经网络(BP)和随机森林(RF)3种方法,分别建立土壤有机质、全氮的估测模型并评价模型的精度与稳定性。【结果】 光谱经不同变换后,与土壤有机质和全氮的相关系数有所提高,且特征波段更为明显,一阶导数与倒数的一阶导数变换优于其他变换,FD-PLSR模型预测有机质精度最高,Rv²、RPD分别为0.89、2.63;(1/R)’-PLSR模型预测土壤全氮精度最高,Rv²、RPD分别为0.83、2.42。【结论】 基于高光谱技术与机器学习模型可以估测博州耕地土壤的有机质与全氮含量。

关键词: 高光谱, 土壤有机质, 土壤全氮, 光谱估测, 偏最小二乘回归, 随机森林