Hyperspectral Estimation Of Soil Organic Matter Content Based On Grey Relational Local Regression Model

Soil organic matter is an important parameter of soil physical and chemical properties,and also an important indicator to measure soil fertility.The traditional soil organic matter determination method is time-consuming,laborious and costly,which cannot meet the needs of precision agriculture.At present,hyperspectral remote sensing technology has become a new technology for soil organic matter monitoring because of its advantages of rich information and real-time and fast.However,due to the diversity and complexity of soil spectral factors,the accuracy of hyperspectral estimation of soil organic matter content needs to be improved.Therefore,this paper takes Zhangqiu District of Jinan City as the research area,and 76 cinnamon soil samples collected as the research object.Based on the obtained spectral data,the grey correlation analysis and statistical analysis method are combined to establish the hyperspectral grey correlation local linear regression model of soil organic matter content.The main research contents and conclusions are as follows:(1)The spectral characteristics of cinnamon soil are analyzed.Firstly,the nine-point weighted moving average method is used to smooth the soil spectral curve,and then the samples are divided into four groups according to the organic matter content.The average spectral reflectance of each group is calculated,and the reflectance curve is plotted and its spectral characteristics are analyzed.The results show that in the range of 350nm-1350nm,the spectral curve has been in an upward trend,the fastest rise in 550nm-850nm,the rest of the band growth slightly;In the range of1450nm-1800nm,the spectral curve increases first and then decreases,and the change is small,the curve is stable;In the range of 1900nm-2100nm,the spectral curve showed a significant upward trend;In the range of 2200nm-2500nm,the spectral curve showed a significant downward trend;The outdoor spectrum has large noise at the absorption valleys of water at wavelengths of 1400nm,1900nm and 2200nm.With the increase of organic matter content in cinnamon soil,soil spectral reflectance decreased.(2)Spectral characteristic factors of soil organic matter were extracted.The logarithm and the first-order differential of square root were used for spectral transformation of soil reflectance,and the correlation coefficient between spectral transformation value and soil organic matter was calculated.The characteristic factors were selected according to the principle of maximum correlation.The results show that the first-order differential transformation of the square root reciprocal and the first-order differential transformation of the logarithmic reciprocal are better.Therefore,the transformation values of the first-order differential transformation of logarithmic reciprocal at 542nm,1507nm,1573nm,1621nm,2106nm and 2288nm and the transformation values of the first-order differential transformation of square root reciprocal at 864 nm were selected as the spectral characteristic factors of soil organic matter,and their correlation coefficients were 0.68,0.70,0.74,0.78,0.82,-0.73 and-0.67,respectively.(3)The hyperspectral grey correlation local linear regression model of soil organic matter content was established.Firstly,the grey correlation degree between the samples to be identified and the known model samples is calculated,and the known model samples are sorted according to the size of the grey correlation.Then the local linear regression model of grey correlation is established by selecting different samples with large grey correlation degree,and the model is optimized by using the minimum error sum comprehensive index(GRN=F×R~2),and compared with the common modeling methods.The results show that the two models based on minimum error and comprehensive index optimization are effective.The determination coefficients R~2of 14 test samples are 0.855 and 0.924,respectively,and the average relative errors are 9.881%and 6.607%,respectively.The determination coefficient R~2of multiple linear regression,BP neural network,support vector machine and classical grey correlation degree were 0.900,0.887,0.864 and 0.531,respectively.The average relative errors were 11.387%,9.262%,12.136%and 17.530%,respectively.Studies have shown that the hyperspectral grey correlation local regression estimation model of soil organic matter content is feasible and effective.