Study On Prediction Of Soil Organicmatter Using Diffuse Reflectance Hyper Spectra

The prediction of soil organic matter content by Hyperspectral data is one of the mostimportant researches in soil remote sensing. In this study, farmland soil in Weiyang District ofXi’an city was taken as the research subject. Hyperspectral prediction modeling on soilorganic matter content and the analysis of its accuracy were performed through the fullspectral data and extracted features, based on field survey sampling, laboratory analysis andspectral testing. The results are as follows.(1)The effects of different spectrade-noising methods were quite different. The effect ofmoving window smoothing method was mainly determined by the size of the smoothingwindow. The effect of wavelet smoothing method depended on the choice of mother wavelet,the setting of wavelet decomposition level and the processing rules in wavelet domain.Inappropriate de-noising method for weak spectral noise signal could cause the loss of usefulinformation, introduce new noise, and reduce the predictive accuracy.(2)The spectral characteristics would be enhanced after Absorb and SNV treatment,which had clear physical meaning and purpose. The prediction model established based onthese data was superior to that based on original reflectance spectra. Differential treatmentcould enlarge the changing information in the original spectrum, but noise information wouldbe amplified as well. Therefore, the appropriate calculation methods and smoothing methodsmust be selected while dealing with spectral data.(3) The partial least squares regression (PLSR) is a good method to establish the spectralanalysis model, the only requirement of which is to ensure the number of principalcomponents. The modeling method could be operated simply and its result was easy tointerpret. Compared with PLSR prediction models established by original reflectance spectramethod and other forms of data transformations, the SNV data conversion method was better,and the effect of K-M transformation were superior to the original spectra.(4)The modeling precision and especially the predicting precision based on supportvector machine regression (SVR) method were better than PLSR method. SVR could obtaingood effect and generalization performance on establishing the nonlinear correlative mappingmodel between full spectrum data and target property. Particle swarm optimization algorithm(PSO), which needed certain experiment and experience as setting parameter, was anexcellent method to choose parameter for SVR method, and possessed better effect and predicting precision than conventional grid search method.(5)Soil organic carbon prediction model created by spectral feature extraction based onthe Gaussian kernel principal component analysis (KPCA) method was better than PLSRmodel and SVR model in predicting effect, while the polynomial kernel principal componentanalysis was inferior to the linear principal component analysis (L-PCA). Choosing the typesand parameters of kernel function were critical when applyingKPCA method.(6)Wavelet analysis (WA), with stability and adoptability, was based on single-spectraldata, and spectral characteristics data could be extracted by WA without knowing thestatistical or geometrical characteristic between spectral data. Low frequencyapproximatecoefficients of the discrete wavelet transform was used as the spectral feature extractiondata,which could improve the predicting results of PLSR modeling.