Spectral Inversion Of Soil Organic Matter Content With Different Levels Of Disturbance

In order to explore the estimation methods of soil organic matter content under different degrees of human disturbance,this paper takes the soil of Fukang City,Xinjiang as the research object,and combines the field measured hyperspectral and multi-spectral images with soil organic matter content to establish the soil organic matter content inversion model.The hyperspectral characteristics of soil organic matter under different disturbance levels,the correlation between soil original hyperspectral reflectance and its different transformed spectral information and soil organic matter content is studied to choose sensitive bands.The soil organic matter content estimation model is established by partial least squares regression and extreme learning machine algorithm,and the optimal model is verified by Kriging interpolation method.At the same time,in order to let the ground measured spectroscopy results be well applied to the soil organic matter estimation of remote sensing images,and improve the accuracy of using the remote sensing data to estimate the soil organic matter content in a large area,the hyperspectral reflectance value is fitted to the multi-spectral band and construct a partial least squares model separately.The main conclusions are as follows:(1)The whole soil organic matter content in the study area is low.The soil organic matter content in the mild interference zone rang is 6.507~15.221 g/kg,and the moderate interference zone varies between 6.182 g/kg and 14.707 g/kg,while the severe interference zone the organic matter content rang is 3.619~13.107 g/kg,and the average values are 10.257 g/kg,8.939 g/kg,and 7.977 g/kg,respectively.As the intensity of the disturbance decreases,the management mode gradually returns to the natural state,the organic matter content gradually increases,and the coefficient of variation gradually decreases.(2)Because of the less organic matter content and the more ground information contained in the field spectral data,combined with the influence of human disturbance,it is relatively difficult to estimate the organic matter content,and the correlation coefficients have not passed the 0.05 level test.(3)Five traditional mathematical transformations of soil original hyperspectral reflectance,including spectral reciprocal,logarithmic,first-order differential,first-order differential of spectral reciprocal,first-order differential of logarithm are carried out.It is found that the correlation between the spectral reflectance value and soil organic matter content in different regions is improved in varying degrees after the treatment of traditional mathematical methods.Five sensitive bands with good correlation are selected to construct partial least squares regression and extreme learning machine models respectively.Through comparison,the best effect of organic matter models in mild,moderate and severe disturbance zones is the extreme learning machine model constructed with first-order differential,first-order differential of spectral reciprocal and first-order differential of logarithm as independent variables.The verification set R~~2 is 0.719,0.691,0.658,and RMSE is 1.134 g/kg,1.519 g/kg,1.879 g/kg,RPD is 1.911,1.817,1.691,respectively.It is shown that the model constructed with first-order differential and first-order differential of spectral reciprocal as independent variables can accurately predict soil organic matter content in mild and moderate interference areas.(4)After continuous wavelet transform,the correlation coefficient between the smoothed original spectrum and soil organic matter content increased greatly.The maximum R~~2 between them is up to 0.48,0.45 and 0.40 in mild,moderate and severe interference areas,respectively.Compared with the original spectrum,the determination coefficient R~~2 is increased by 0.33,0.28 and 0.37,respectively.Partial Least Square Regression and Limit Learning Machine models are constructed in five sensitive bands with high correlation coefficients.The prediction accuracy of the latter model is better.The accuracy of the model in the mild,moderate and severe interference regions is 0.876,0.791,and 0.729,respectively.RMSE is 0.984g/kg,1.313g/kg,and 1.881g/kg,respectively.And the RPD is 2.492,2.369,2.264,respectively.That is,the extreme learning machine model can accurately predict the soil organic matter content in different interference zones.(5)In order to further explore the estimation method of soil organic matter content under different degrees of human disturbance,the smoothed 1~8 layer decomposition and reconstruction of the smoothed original spectrum are carried out.After decomposing the original spectra of the three regions at different scales,the L3and L4 models are the best in general.Compared with L0 models,the verification set R~~2 increases by 0.06,0.04 and 0.05 respectively,which proves that L3 and L4 retain the most effective spectral information while denoising.The gray correlation between soil spectral reflectance and organic matter content was analyzed by grey correlation,and the extreme learning machine model is constructed.The model accuracy R~~2 is0.825,0.787 and 0.729,respectively,and the RMSE is 0.980g/kg,1.387g/kg,and1.982g/kg,respectively.RPD reaches 2.382,2.204,2.139,respectively.(6)Corresponding to the bands of Landsat image,the hyperspectral data is fitted into multi-spectral image band information.The correlation coefficients of the mild,moderate and severe interference regions are 0.981,0.977 and 0.968,respectively,which have good consistency.The accuracy of the partial least squares regression model constructed by multispectral fitting using hyperspectral fitting is0.833,0.722,0.66,which is generally higher than that of multispectral band models.After adjusting the multi-spectral band by the ratio correction method,the R~~2 of the mild,moderate and severe interference areas increased by 0.094,0.123,0.146,respectively,and the RPD increased by 0.398,0.472 and 0.448,respectively.The RPD is greater than 1.8,achieving a good prediction level.