Prediction And Spatial Distribution Of Soil Organic Matter Using Visible Near-infrared Spectroscopy In Weibei Plain

Soil organic matter?SOM?is an important nutrient source for plant growth.It can also adjust soil structure to improve its physical and chemical properties.It is one of the important indicators for measuring soil quality.Traditionally,the monitoring of regional soil organic matter content requires a large number of sample points and laboratory analysis.It is laborious and the experimental results are not repeatable.At the same time,it is difficult to achieve multi-temporal dynamic monitoring in large areas,and the application of this technique is limited.With the development of remote sensing technology,the acquisition of soil organic matter content information is not limited by time and space,and the dynamic change information of soil organic matter can be obtained.The quantitative of soil remote sensing technology based on spectral reflectance characteristics of ground objects has been widely accepted and applied.The spectral prediction model can quickly obtain information on soil organic matter content in the large regions.The Weibei Plain is the main research area.Based on the indoor sample layout planning in the early time,the field hyperspectral data collection and soil physical and chemical composition are analyzed on the field soil samples,and the soil organic matter content is explored by combining the remote sensing image data.The intrinsic relationship between the value and the spectral reflectance data were determined,and finally the remote sensing prediction model is established to estimate digital mapping of soil organic matter and the dynamic monitoring of land quality on the regional scale.This research mainly includes the following procedures and results:?1?By analyzing spectral data of the soil samplings and the reflectivity characteristics of the corresponding Landsat-8_OLI image,it is concluded that the sensitive bands closely related to SOM are mainly at 345,491,594,627,718,723,748,810,873,889,931,981 and 1075 nm.The correlation between soil organic matter and image spectral is higher.There is a significant difference in soil reflectance at different points in the visible-near-infrared?Vis-NIR?range.The higher soil organic matter content accompanies the lower spectral reflectance of the soil.The position of the spectral reflection peak in the range of 600nm-800nm is related to soil degradation.These subtle changes are crucial for establishing the SOM inversion model.?2?Using the field hyperspectral data on the ground can effectively predict the soil organic matter content,and the SVM model imported into the constructed spectral index has the highest prediction accuracy and stability.The correlation between the original hyperspectral data and SOM is weak,and it is difficult to judge the characteristic band.The spectral first-order inverse differential transforms processing can highlight the characteristic of the hyperspectral data for soil information.After the transformation,the correlation coefficient of SOM alternates with positive.The absolute value of the coefficient is increased to facilitate the selection of the feature band.After completing the selection of the characteristic band,Principal Component Analysis is used to convert the band spectral information into six linear uncorrelated spectral principal components.Based on this,the difference spectral index?DI₁ and DI₂?is constructed,and then the spectrum indices are extracted.The SOM measured values are used as independent variable and dependent variable to establish multiple linear regression?MLR?model,error reverse transmission multilayer perceptual neural network regression?BP?model and support vector machine regression?SVM?model,and the inversion model were established.The target requirement is that the SVM model has the best prediction accuracy,Coefficient of Determination?R²?and Root Mean Square Error?RMSE?are 0.803 and 0.223,respectively,and the R² and RMSE of the BP model are 0.764 and 0.161 respectively.The prediction accuracy and stability of the MLR model are unsatisfactory,R² and RMSE are 0.704 and 0.151,respectively.?3?The mathematical transformation of Landsat-8_OLI original spectral data can eliminate the influence of background information such as moisture.The BP model using spectral transformation?CD?to construct the spectral index has the best prediction result,and the soil organic matter in the study area can be obtained by spectra inversion.The first-order differential transformation can improve the accuracy and stability of the multiple linear regression?MLR?model.The R² and RMES of the model are 0.585 and 0.268,respectively,and the R² is increased by 0.11 compared with the untransformed.For the BP neural network model,the reciprocal transform?CD?can highlight the spectral characteristics of each band.The transformed model operating parameters R² and RMSE are 0.65 and 0.250,respectively,were higher than the original band data and the rest of the mathematical transformation,and the RPD value is 1.24,the model is acceptable but there is still details for further improvement;when using the image spectrum and its mathematical transformation data to estimate the SOM concentrations,the support vector machine regression?SVM?model cannot correctly seek the local minimum due to the large sample over-fitting problem.The prediction accuracy is lower than other models,R² and RMSE are 0.272 and 0.185,respectively.?4?The spatial distribution of soil organic matter in Weibei plain shows the overall trend of soil organic matter content decreasing from south to north.Based on MLR,BP,SVM and Ordinary Krige?OK?prediction models,the spatial variation trend of soil organic matter can be well fitted.The soil organic matter content in the eastern and western regions of the study area is high.The middle area of the Bailang River and the Wei River is affected by seawater inversion,and the soil has been degraded of organic matter content.The prediction result of BP model is the best,the spatial variation of predicted SOM content is more obvious,and the prediction model is more stable and accurate,while MLR Model and SVM regressions do not handle eigenvalues well,resulting in soil organic matter contents that vary too much or too little throughout the region.Ordinary Krige depend on the content of an unknown sample based on the semi-variogram model of regional variables,which are greatly influenced by artificial sample.This study can be analyzed separately or as a powerful tool for verification of remote sensing inversion model,the inversion model can make good use of the association law of regional variation variables in geostatistical methods,and provide basis for the improvement of inversion models.