Estimation Of Soil Organic Matter,Total Carbon And Total Nitrogen By Hyperspectral Remote Sensing In Qinghai Province

Soil organic matter,total carbon,and total nitrogen content play an important role in crop growth.Rapid access to information on organic matter,total carbon and total nitrogen in soil is of great significance for the scientific management of agricultural production and the development of precision agriculture.The traditional measurement method mainly relies on the method of laboratory chemical analysis.The operation process is complicated,time-consuming,and the measurement component is high,which is not conducive to large-area acquisition measurement and accurate grasp of soil information.Visible-near-infrared spectroscopy?also known as visible-near-short-wave infrared spectroscopy?Vis-NIR-SWIR??has the advantages of low cost,no pollution,and rapid real-time detection,and has been widely used in soil property detection.In this paper,the Sanjiangyuan region and the Hangshui river basin in Qinghai Province were used as research areas,and the soil samples collected in the field were measured for physical and chemical properties and soil spectra.Based on this,the soil spectral data were preprocessed and based on fuzzy K-means clustering.The soil spectrum was classified and characterized.Finally,the hyperspectral estimation of soil organic matter,total carbon and total nitrogen content in the two study areas was carried out based on partial least squares regression?PLSR?and random forest?RF?.The main conclusions are as follows:?1?The spectral reflectance of soil shows as a whole that the higher the content of soil organic matter is,the lower the spectral reflectance is;and the reflectance varies greatly from 400 to 1350nm.The higher the content of organic matter is,the more concave the reflectance curve is.With the increase of organic matter content,an obvious absorption valley appears in the spectral curve of the envelope removed in the band of 400¹350nm.?2?Envelope removal?or continuum removal?can highlight the absorption and reflection characteristics of the soil spectrum and facilitate the classification of soil spectra.Based on the fuzzy K-means clustering of soil envelope removal,the soil spectral classification of all soils?two study areas?was carried out.The classification results showed that the soil spectral classification of this area was greatly affected by soil organic matter content.?3?The spectral characteristics of the soil after first-order differential and envelope removal treatment of the soil spectrum are more prominent.The full-band partial least squares regression modeling of the original spectrum,first-order differential,and envelope removal process shows that the first-order differential and envelope removal can reduce the principal component of the model and simplify the model structure.?4?The stable competitive adaptive re-weighting sampling selects the feature band modeling and the full-band modeling.Compared with the full-band modeling accuracy,the partial-lean-square regression and random forest estimation models have no improvement in the overall band modeling accuracy.However,feature band modeling can simplify the model and increase the computation rate.The characteristic bands extracted after different pretreatment of the soil spectrum are also different.?5?The modeling set selected by the concentration gradient method and the Kennard-Stone?KS?method was used to verify the statistical characteristics of the physicochemical properties,and a partial least squares regression model was established.The modeling of the KS method in the Huangshui river Basin is not high,and the R²_cvv is large,the R²_val is small,the model is over-fitting,and the models of the two modeling sample selection methods of Sanjiangyuan region have good estimation accuracy.The organic matter,total carbon and total nitrogen content of Sanjiang yuan region are higher than those of the Huangshui river basin,indicating that the model set selected by the concentration gradient method is more representative.When the soil organic matter content is low,the KS method selection modeling set is not representative.?6?In the estimation of soil organic matter,total carbon and total nitrogen content,the estimation accuracy of random forest is higher than that of partial least squares regression model.This may be because soil organic matter,total carbon,total nitrogen content and soil spectrum are not simple.Linear relations,partial least squares regression modeling has certain limitations.?7?The accuracy of the soil estimation model based on fuzzy K-means clustering analysis is not improved compared with the accuracy of the global soil estimation model and the accuracy of some types of estimation models is less than the accuracy of the global estimation model.It may be due to the small difference in soil in this area,the influence of organic matter on the soil is large,and the number of samples after classification is small.