Estimation Of Soil Organic Matter Content Based On Vis-NIR Spectroscopy And Regression Techniques

Soil Organic Matter is an important index of soil fertility and soil quality.Although the measurement accuracy of soil organic matter content using traditional traditional chemical analysis methods is better,but the analysis process has a long period,high cost,and can only detect one item at a time,and it has certain pollution to the environment,so it is difficult to use it on a large scale.In view of the importance of soil organic matter content,there is an urgent need for a rapid,simple,low-cost,environmentally-friendly and non-polluting detection method.Visible near infrared spectral technology can quickly obtain the information of the reflection spectrum on the soil surface,and because of its extremely high spectral resolution,it can produce a complete continuous spectrum curve that can reflect the information of the ground features,which makes visible-near-infrared spectroscopy analytical techniques show great potential in the predictive analysis of soil physical and chemical composition.In this paper,a total of 428 soil surface samples from the HuangShui River Basin of Qinghai Province were collected for light measurement and physicochemical analysis indoor.The spectral range is 350~2500 nm.After spectral pretreatment and characteristic wavelength selection,Partial least squares(PLSR),support vector machine(SVM)and random forest(RF)regression model are established to investigate the feasibility of using Vis-NIR spectral analysis and regression techniques to quickly estimate soil organic matter content and it provides new ideas for digital soil mapping and land quality evaluation.The main research contents and results are as follows:(1)Based on the original spectrum and the pretreatment spectrum,select the modeled samples by the Rank-Select method and the Kennard-Stone method in a ratio of 2:1 to construct PLSR model.The determination coefficient(R~2)and relative analysis error(RPD)obtained by the PLSR model based on the Rank-Select method are higher than the Kennard-Stone method in two spectral modes.The component content of the calibration set sample partitioned by the concentration gradient method covers the component content of the validation set sample,avoiding too many“special”samples to be divided into calibration sets,so that the established model can better predict the unknown sample.(2)Comparing the accuracy of the PLSR model for 12 soil spectral data preprocessing algorithms.The optimal preprocessing algorithms are multi-dimensional scatter correction(MSC),median filter(MF)and first derivative(1st Der)use together.Based on the PLSR model established by the MSC-MF-1st Der algorithm,R~2 is 0.84and RPD is 2.5,which is superior to the accuracy of the original spectral PLSR model and has better prediction ability.(3)Application of the Stabilization Competitive Adaptive Reweighting Algorithm(sCARS),Successive Projection Algorithm(SPA),Genetic Algorithm(GA),Iterative Retention Effective Information(IRIV)and sCARS-SPA Algorithms from the Raw Spectra and MSC-MF-1st Der Pretreatment Spectroscopy to extraction of Characteristic Wavelength Variables.The feature variables selected by the five variable select methods are mainly distributed in the near-infrared spectral region.(4)After the original spectra were pretreated by MSC-MF-1st Der,the accuracy of the PLSR and RF models for the six spectral variables(full band and five characteristic bands)was higher than that of the original spectrum.In addition to the SPA characteristic variables,the accuracy of SVM models for the other five spectral variable is higher than that of the original spectrum.Preprocessing the original spectrum can improve the accuracy of the spectral model.(5)After spectral preprocessing,PLSR model,SVM and RF models were constructed based on the full-band variables and characteristic band variables.The PLSR and SVM models were modeled using characteristic band variables.The number of variables used to construct the model was significantly reduced,and the prediction ability of the model was somewhat higher than that of the full band.The root mean square error of R~2obtained from the six spectral variable validation set of RF model was 0.00232,and R~2 basically had no obvious change.Therefore,the RF model adopts characteristic wave band modeling,which does not help the improvement of the model accuracy,but the variables numbers of its construction model are significantly reduced,which greatly improves the modeling efficiency.The characteristic variable selection from the full-spectrum can greatly reduce the complexity of the model while ensuring the accuracy of the model.The order of the five algorithms for simplifying model capabilities is as follows:sCARS>IRVI>GA>sCARS-SPA>SPA.(6)Among the three regression models,the RF model has the best forecasting effect and is superior to the SVM model and the PLSR model.The MSC-MF-1st Der-sCARS-RF has the best prediction effect,with 51 characteristic variables,accounting for only 2.55%of the whole band,R~2 is 0.958,and an RPD is 4.5 of the validation set.This method can well predict soil organic matter content.The relationship between soil organic matter and spectrum is not a simple linear relationship,so the PLSR model shows some limitations,and SVM and RF can better solve the complex nonlinear relationship between independent variables and dependent variables.However,the SVM model tends to cause serious deviation estimation due to higher spectral noise and the model accuracy decreases.The RF model has better anti-noise ability,which makes the established model more accurate and has better robustness.