Estimation Model Of Soil Nutrient Content Based On Hyperspectral Technology

Nitrogen,phosphorus andpotassium in soil are vital nutrients for crop growth and evelopment,which are important indexes for evaluation of soil fertility and crop growth status.Therefore,rapid estimation of soil nutrient content is indispensable for monitoring crop of growth and utilizing land resources reasonably.The traditional measuring method of nutrient elements exists serious problems such as time consuming and laborious.Hyperspectral remote sensing as a hot quantitative remote sensing technology,taking advantage of continuous spectral channels record the complete spectral information of soil,rapid estimation of soil nutrient content can be achieved by establishing an estimation model between the content of nutrient lements and soil characteristic spectrum,which provides an efficient and convenient means to obtain soil nutrient content data and plays a crucial part in soil quality evaluation and precision agriculture management.This paper took 92 soil samples and ground hyperspectral data as the data source,and took Hengdaohe Town,Liaoyuan City,Jilin Province as the research area.After the pretreatment of soil spectral data such as smoothing and denoising,breakpoint correction,spectral characteristic transformation and resampling,The correlation between soil nitrogen,phosphorus and potassium contents and spectral characteristic transformation such as spectral reflectance,absorbance,continuum removal and first order differential was analyzed.Through screening various nutrient elements characteristics of sensitive wave bands by significant tests,and using competitive adaptive weight weighted algorithm for optimal transform form of dimension reduction of sensitive wave bands,which were selected multilayer perceptron neural network,support vector machine and gradient promotion tree,which are regarded as three kinds of machine learning algorithm to build models for predicting the soil nutrient elements in the research area respectively,and the accuracy of the model estimation was compared and analyzed by calculating the model determination coefficient,root mean square error and other precision indexes.Through the analysis and research,the main results of the research are as follows:The overall tendency of the spectral curve of soil samples in the research area is almost identical,and the spectral reflectance reduces gradually with the increase of wavelength.With the increase of the content of nitrogen,spectral reflectance declines,but the correlation law with the content of phosphorus and potassium is not obvious.Both continuum removal and first-order differential treatment can highlight the details of spectrum and enhance correlation.The content of nitrogen and potassium possess the highest correlation with the first-order differential of spectral absorbance,while the content of phosphorus owns the highest correlation with the first-order differential of reflectance.Among the three above-mentioned models,the results of gradient ascending tree regression modeling are much better than multilayer perceptron neural network and support vector machine,and the accuracy of the estimation of nitrogen is significantly higher than that of phosphorus and potassium.The estimation model is built by utilizing spectral data after dimensionality reduction of competitive adaptive reweighting algorithm,which the ability and stability of prediction are promoted to a certain extent.