A Study On Estimation Of Total Soil Nitrogen And Organic Matter Content Based On The Near-Infrared Reflectance

As important indicators of soil fertility, soil organic matter and total nitrogen are significant in agricultural production. Thus, how to determinate soil organic matter and total nitrogen content rapidly and accurately has been a hot issue. Near-infrared NIR (Near Infrared Reflectance, NIR) technology, with its advantages of quick and low consumption, simultaneous determination of multicomponent, can meet the requirements of large-scale soil information rapidly and accurately grasp.In this paper, samples of different soil types are collected in Qiqihar and Suihua from Songnen Plain. Soil NIR spectral estimation model is built referring to different spectral information disposing methods and different modeling methods, using the combination of internal cross validation and external validation, choosing determination coefficient and root-mean-square error as indicators to evaluate the accuracy of different estimation models.First of all, the study analyzes the effect of different preprocessing methods on soil spectra1information. It is found that using the three wavelet decomposition on the basis of the db5wa velet to process the near-infrared spectral data could de-noise on the original spectrum and bett er preserve the spectral information than moving window smoothing method and minimum squ ares fitting method. Then it uses the partial least squares combined with different preprocessing methods to establish the inversion model of soil organic matter. By comparing the effect of different combinations on the model, it’s concluded that the method of the wavelet transform combined with second derivative and multiplicative scatter correction is the best preprocessing method and its decision coefficient is up to0.884.Secondly, compare the soil properties model accuracy of the soil spectral estimation with different established methods. In this study, PLS, WNN and LSSVM are used to establish soil characteristics estimation model by near-infrared spectral. The results show that:the PLS model is the simplest, lower accuracy, but very good promotional value. The LSSVM model has the higher accuracy, but the lack of stability. The WNN model is the best with very high R2and lower RMSE which is respective0.974and0.0139, external validation R2and RMSE which is respective0.9627and0.0130; cross-validation R2and RMSE of WNN soil organic content estimation model is respective0.990and0.2097, validation R2and RMSE of WNN soil organic content estimation model is respective0.9827and0.2114, which the high relation of soil organic and NIR.Meanwhile, the choice of the nodes number in the hidden layer is particularly important in the process of establishing WNN model. The small numbers lead to the models failing to meet requirements, the model is more complex with large numbers and may appear over-fitting. The model is the best which the number of nodes in the hidden layer is5. You must determine the penalty factors and kernel parameters before LSSVM algorithm when modeling, otherwise the accuracy of the model will be impacted seriously. Selecting parameters by ten-fold cross-validation method and finding that which σ2and C is respective11.6888and24.9289. Soil organic matter estimation model is optimal which σ2and C is respective11.7878and26.7735.