| The overall situation of soil nutrient content is of great significance to the growth of crops,and the estimation of available nutrient content in farmland soil is a key link in the evaluation process of crop growth and development.The development of hyperspectral remote sensing technology provides an effective means for the dynamic monitoring of regional organic matter and soil nutrients.Compared with traditional methods,such as time-consuming,labor-intensive and complicated measurement,hyperspectral remote sensing technology has the characteristics of wide monitoring range,dynamic monitoring and high spatial resolution.Therefore,in order to explore the relationship between soil nutrient content prediction and hyperspectral data,this paper takes the research and experimental base of Jilin Agricultural University in Changchun as the research area,collects soil samples in the research area,and measures the nutrient content and spectral data in the soil samples in the laboratory.Based on the use of Arc GIS10.2,EXCEL2018,Origin2018,Matlab2022 a and the mathematical transformation of original spectral data,the soil in the study area was finally evaluated,and the support vector machine model and BP neural network model of soil organic matter,total nitrogen,total phosphorus and available phosphorus content were established.By comparison,the best inversion model of soil nutrient content was obtained.The main research results are as follows:(1)Five-point sampling method was used to collect soil samples from the land in the study area,and 163 sampling points were set up.The indoor nutrient content and spectral data of the collected soil samples showed that the contents of organic matter,total nitrogen and total phosphorus in the soil samples were positively correlated;Moreover,the spatial interpolation results show that the overall distribution gap is large,and the soil total nitrogen in the whole study area belongs to a very rich level;The total phosphorus content of soil belongs to the middle and lower grades;The contents of organic matter and available phosphorus belong to rich grades.(2)By measuring the spectral reflectance of soil samples in the study area,it is known that the spectral reflectance of sampling points will decrease with the increase of nutrient content.The correlation coefficients of organic matter,total nitrogen,total phosphorus and available phosphorus are low in the original spectrum and four kinds of mathematical change spectra,but the correlation coefficient of spectral data has increased greatly after continuous wavelet transform,which shows that continuous wavelet transform can not only improve the correlation coefficient between soil nutrients and spectral data,but also accurately predict the soil nutrient content.(3)By comparing the results of each nutrient support vector machine model and BP neural network,we can know that the best prediction models of soil organic matter,total nitrogen,total phosphorus and available phosphorus are all BP neural network models established after continuous wavelet changes,in which the best prediction model of organic matter is BP-CWT-Ln R,the best prediction model of soil total nitrogen is BP-CWT-R "and the best prediction model of soil total phosphorus is BPCWT-1/".The best prediction model of soil available phosphorus is BP-CWT-1/R,and the minimum absolute coefficient R2 used to verify the accuracy of the model is0.9623,and the verification result is close to 1,which shows that the prediction effect is excellent.(4)In the model of support vector machine,the prediction effect of RBF kernel function is higher than polynomial kernel function and linear kernel function,but its stability is not as good as that of BP neural network model,and the highest prediction result of soil total phosphorus by support vector machine model is R2=0.4167 after the verification of absolute coefficient,which shows that support vector machine model is not suitable for predicting soil total phosphorus content. |
PDF Full Text Request