Estimating Cultivated Soil Fertility Properties From Visible-near Infrared Reflectance Spectroscopy (VNIR)

Soil nutrient status,soil texture and p H are important indicators of soil fertility,which are important information bases for accurate fertilization.With the development of visible-NIR hyperspectral technology,the laboratory measurement-based soil visible-NIR hyperspectral technology will gradually replace the traditional laboratory chemical analysis-based fertility attribute detection methods with the advantages of fast analysis speed,simple operation and low detection cost,providing a new way for soil fertility attribute content detection.However,the soil spectrum measured under controlled laboratory conditions is a kind of soil spectral curve in an ideal state,and the pre-treatment work such as air-drying,grinding and sieving of soil samples is still needed before the spectral measurement,which requires a lot of labor and material resources.Compared with laboratory visible-NIR spectroscopy,field in situ visible-NIR spectroscopy is more effective in improving the collection efficiency of soil spectra and obtaining the most current spectral data,so it has received more attention.However,in situ field spectra are easily disturbed by environmental factors in the acquisition process,which leads to the accuracy of soil property content prediction models established by them being much lower than that of laboratory spectra.How to improve the accuracy of soil property content estimation by field in situ visible-NIR spectroscopy,so that it can meet the precision requirements in precision agriculture while obtaining soil information quickly,is a hot spot and difficult area of soil hyperspectral technology research in recent years.Based on the above considerations,this paper takes Huangshui watershed of Qinghai province as the research area,based on a total of 220 samples of surface layer（0-20 cm）of cultivated soil collected in 2015,2016,and 2017 as the research object,and uses soil visible-near infrared reflectance spectroscopy（VNIR）Spectroscopy（VNIR or vis-NIR）combined with chemometric methods to explore the ability of field in situ spectroscopy and corresponding laboratory spectra to estimate the content of 12 fertility attributes such as total carbon（TC）,organic matter（OM）,and total nitrogen（TN）in soil at the watershed scale.On this basis,in order to improve the accuracy of field in situ spectra for estimating soil fertility attributes,a moisture impact correction algorithm was introduced to eliminate soil moisture noise in field in situ soil spectra and compare the correction effect of different algorithms on soil moisture impact;meanwhile,field and laboratory spectra were combined.The aim is to further improve the prediction accuracy of field in situ soil spectra for soil fertility attributes,and provide new ideas and methods to achieve rapid,accurate and nondestructive determination of soil attribute content.The main research contents and conclusions of this thesis are as follows:（1）Characteristics of field in situ and laboratory soil spectral curves and estimation resultsThe distribution characteristics of field in situ spectra and laboratory spectral reflectance curves of different soil samples are basically the same,but the field in situ spectral reflectance is lower than the laboratory spectral reflectance curve in the whole waveband range.The results of estimating 12 soil fertility attributes based on field in situ spectra and their corresponding laboratory spectra showed that the contents of five soil fertility attributes,namely,total carbon（TC）,organic matter（OM）,total nitrogen（TN）,alkaline nitrogen（AN）and p H,could be estimated by both laboratory spectra and field in situ spectra,and the laboratory estimation results were higher than the accuracy of field spectra.TK,AK,TP,AP,and clay,silt,and sand in soil texture could not be estimated by laboratory and field in situ spectroscopy.（2）The effective combination of spectral mathematical transformation,feature band selection and regression model can improve the accuracy of soil property content estimationThe SG first-order differential transformation can effectively enhance the field spectral features,and the stability competition adaptive reweighted sampling method algorithm has the best effect on the extraction of the characteristic bands of soil fertility attribute content.The random forest model had the highest estimation accuracy for the five estimable soil fertility attributes,namely,total carbon（TC）,organic matter（OM）,total nitrogen（TN）,alkaline nitrogen（AN）content and soil p H,and was superior to the partial least squares regression model and the support vector machine model,and for total carbon（TC）,organic matter（OM）,total nitrogen（TN）content and p H,the best estimation model was The s CARS-RF model with SG first-order differentiation had the smallest root mean square error in the validation set among the six models,with RPD values above 4.00 RPD_TC-Field=4.01,RPD_OM-Field=4.72,RPD_TN-Field=4.22,RPD_AN-Feild=4.68,RPD_{p H-Feild}=4.89;RPD_TC-Lab=6.66,RPD_OM-Lab=6.34;RPD_TN-Lab=7.39;RPD_AN-Lab=6.11;RPD_{p H-Lab}=5.00).（3）Moisture removal algorithm can improve the accuracy of field spectral estimation to some extentThe correlation between the field in situ spectra corrected by the spectral direct conversion method（DS）and the orthogonal signal correction algorithm（OSC）and the content of soil fertility attributes improved significantly compared with the field in situ spectra among the four corrected spectra,with the absolute values of the maximum correlation coefficients ranging from 0.14 to 0.82;among them,the spectral direct conversion method（DS）showed a better correction effect on the field in situ spectra,and the five soil fertility All the field in situ spectra of the five soil fertility attributes were corrected by DS with the highest model estimation accuracy,and the RPD could reach above 1.70.（4）Combined laboratory spectra and field in situ spectra combined with full principal component stepwise regression models are feasible for estimating soil organic matter contentThe principal component stepwise regression model combined with the field in situ spectra could not estimate the soil organic matter content,but the principal component stepwise regression model（PCSR）built from the field spectra corrected by the direct spectral transformation（DS）method achieved a good estimation of soil organic matter content with an RPD up to 2.23.In addition,using the principal components of the laboratory spectra extracted from the principal component stepwise regression model as a guide The four principal component stepwise regression models constructed using correlation analysis to select the principal components of field-corrected spectra were able to roughly or well estimate soil organic matter content with RPDs ranging from 1.54 to 2.28,among which the principal component stepwise regression model constructed by combining the principal components of laboratory spectra with field DS-corrected spectra had the highest accuracy with an RPD of 2.28.