Quantitative Inversion Of Total Nitrogen In Cultivated Soil In Southern Xinjiang Based On Visible And Near-infrared Spectroscopy

hyperspectral estimation models of soil total nitrogen were constructed to rapidly and accurately monitor soil total nitrogen content in farmland,thus providing new technology and method for judging crop growth and evaluating land quality.The main cultivated soil types in Southern Xinjiang were studied.The total nitrogen content and spectral reflectance of different cultivated soils were measured.In this paper,we combined the spectral reflectance（R）and its four transformation forms with the soil total nitrogen content,and the total nitrogen content estimation models of the whole region and the region were established by using three modeling methods:partial least squares regression（PLSR）,support vector machine regression（SVR）and random forest regression（RF）.Spectral estimation models were established for six soil types including silt and yellow mud,yellow tide soil,irrigation and siltation soil,lime-lime mud,paddy soil and saline soil,and their accuracy was verified.The results indicated that:（1）The R² and RPD of the optimal RF model were 0.86 and 3.52,the R² and RPD of the SVR optimal model were 0.75 and 1.97,and the R² and RPD of the optimal PLSR model were 0.73 and 1.82,respectively.It can be seen that the prediction ability of the three models was RF>SVM>PLSR.In addition to the first-order differential transformation,other spectral transformation methods improved the accuracy of the model in different level.RF modeling based on spectral reflectance and its four transformation forms had high accuracy,while PLSR and SVM modeling had relatively low accuracy.The stability of the whole region model was higher than that of the partition model,and the difference of the partition model was obvious and the stability was poor.In general,RF model showed stable prediction ability and good applicability for large sample data,and could accurately estimate soil total nitrogen content.However,PLSR and SVM models could only estimate soil total nitrogen content roughly in the study area.（2）The spectral reflectance of silt and yellow mud was the highest,followed by livid silt,yellow tide soil,irrigation and siltation soil,saline soil and paddy soil.The correlation analysis between total nitrogen content and spectral reflectance of six tillage soil types showed that the maximum positive correlation coefficient was 0.57 at the band of 1820 nm.The maximum negative correlation coefficient was-0.55 at583 nm band.After continuous removal of reflectance,the maximum negative correlation coefficient was-0.75 at 627 nm band,and the maximum positive correlation coefficient was 0.68 at 974 nm band.The model accuracy of the irrigated silt and silt yellow mud was higher than the model accuracy of the full sample.The RF model R² of the irrigated silt was 0.87 and RMSE was 0.05 g kg^-1;the verification accuracy R² was 0.84,RMSE was 0.07 g kg^-1,and RPD was 3.56.（3）When the reflectance data was divided into three categories,it could better reflect the spectral characteristics of various soil samples.The spectral characteristics of all kinds of soil samples could also be better reflected when the removal continuum data were divided into five categories.The reflectance of OR-A,OR-B and OR-C all showed an upward trend,and the reflectance of OR-A was higher than that of OR-B and OR-C.After the continuum removal treatment,the characteristic bands of the CR-I,CR-II,CR-III,CR-IV and CR-V5 spectral curves were basically the same,focusing on the 500nm and 1900nm bands.The characteristic bands of CR-Ⅳand CR-Ⅴwere more obvious.The Rp²,RMSE and RPD of the whole region reflectance model were 0.83,0.08 and 3.24,respectively,which were higher than those of the continuum removal spectrum.In spectral reflectance classification,OR-A had the highest modeling accuracy,with model validation accuracy Rp²,RMSE and RPD of 0.87,0.06 and 3.32,respectively.In the classification of continuum removal spectrum,the modeling accuracy of CR-II was the highest,and the model validation accuracy Rp²,RMSE and RPD were 0.74,0.13 and 3.16,respectively.