Quantitative Hyperspectral Inversion Of Organic Matter In Black Soil Based On Statistical Model

Black soil is an important non-renewable resource.As a major grain-producing area in China,the quality of cultivated land in black soil has a direct impact on China's food security,as well as on economic development and ecological environment balance.The amount of organic matter in soil is an important index of soil quality.The traditional method of obtaining soil organic matter content is usually to collect soil samples in the field and use chemical methods to measure them in the laboratory.This method has high accuracy and strong accuracy,but it is laborious,takes a long time and requires a large amount of basic work,and it is unable to realize the rapid determination of soil organic matter content in regional area,and has great limitations.Hyperspectral remote sensing technology provides a new perspective for the estimation of organic matter content,and while obtaining the data,it is not destructive to soil and meets the development requirements of modern agriculture.The research object of this paper is black soil in northeast China.70 soil samples were collected to obtain laboratory spectral data.Firstly,the breakpoint correction,smooth denoising and 15 different forms of spectral transformation were carried out on the soil spectral curve,and the characteristics of the soil spectral curve were analyzed.Then,the correlation between spectral reflectance and soil organic matter content was analyzed,and the explanatory variables of the inversion organic matter content were determined by combining the step method,and the multicollinearity of the data was studied.Then,based on statistical theory and spectral information,multiple linear regression,partial least squares regression and ridge regression models were established to estimate soil organic matter content,and the results of modeling were compared and analyzed.Finally,Based on BP neural network,the nonlinear model is preliminarily explored.The main conclusions are as follows:1)the reflectivity of spectrum curve decreases with the increase of organic matter content,increases with the increase of the wavelength,various forms of spectral transformation effectively highlight the absorption characteristics of soil spectrum,and improve the correlation coefficient between the spectral reflectance and organic matter content,the first-order differential transformation of the logarithm of the original spectral reflectance has significant effect on improving the correlation between them at 1280 nm,the correlation coefficient is 0.7588.2)in the establishment of partial least squares regression model,root-mean-square error of training set and verification set samples is taken into account comprehensively to determine the number of components involved in the modeling,which improves the reliability of selecting the number of components.The model established by the partial least squares regression method is better than the multiple linear regression,and the modeling effect is the best in the form of logarithmic second-order differential.The correction determination coefficients of the samples of the modeling set and the verification set respectively reach 0.9488 and 0.7864,and the prediction residual exceeds 2.The model has a good prediction ability.3)in the process of establishing ridge regression model,the determination of ridge parameters is a crucial link.In this study,the selection method of ridge parameters was improved,and the ridge trace method,residual sum of squares method and variance enlargement factor method are combined to determine the optimal ridge parameters.It avoids being too subjective in the screening process of ridge parameters and makes the selected ridge parameters more reliable.on the basis of improving the ridge parameter selection method the ridge regression model was established,the precision of the ridge regression model is superior to the former two models,the stability of the model and forecasting ability are improved,the sample points near the 1:1 line more concentrated,The correction determination coefficient of the model training set samples established in the form of logarithmic second derivative of the reciprocal of reflectivity reached 0.9704,and the prediction residual error is 3.1810,it is the optimal model in all linear modeling methods under all spectral transformations.4)The analysis showed that there was serious multicollinearity among the data.For the data in this study,the partial least squares regression and ridge regression modeling methods effectively overcame the multicollinearity problem.5)The nonlinear model established by BP neural network has a good fitting effect on training data,the model not only had high stability,but also achieved higher predictive ability,the training set and validation set sample correction decision coefficient reached 0.9862 and 0.9783 respectively,the comprehensive of evaluation indexes of the model,the model established by BP neural network has the best effect.