Predictionof Fertility Factors In Paddy Soil Using VNIR And MIR Spectroscopy

Soil is variable in spatial distribution and its type is multiple. Therefore, accurate and rapid obtainment of soil properties contributes greatly to digital soil mapping and updating in large scales, while it is also the foundationof environmental problem solving and the ganrantee of modern agriculture implementing. As spectral techniquesboom in the recent two decades, these techniques, especially visable near infrared (VNIR) spectroscopy and mid-infrared spectroscopy, are widely applied in soil properties prediction. Soil VNIR and MIR spectrum are comprehensive response of soil organic matter, mineral, water, particle size and colour, then they can be used to predicte soil properties.There are few research onprediction of soil properties using MIR spectroscopy in China, and no report about comparation prediction accuracy of soil properties using VNIR, MIR and VNIR-MIR in paddy soil is found. In this study, we focus on paddy soil from Zhejiang Province. Spectral characteristic of paddy soil were analysised in VNIR and MIR regions. We are aimming at comparing the prediction accuracy of six soil fertility factors such as organic matter (OM), total nitrogen (TN), available nitrogen (AN), available phosphorus (AP), available potassium (AK) and pH using differnet spectral preprocessing methods, calibration selection methods and chemometric modelsusing VNIR, MIR and VNIR-MIR. Important calibration bands in VNIR and MIR regions were selected by variable selection method from best prediction models, then multiple linear regression models were built by these calibration bands. We hope this study could do contribution to prediction soil fertility factors in paddy soil using VNIR and MIR techniques scientificially, to development and application of soil fertility factors prediction spectromenter, and to fertilization and precise management in paddy soil.The main research routes and results in this study are as follows:1. Comparation of prediction accuracy of soil fertility factors using VNIR and MIR spectroscopy.Valibration datasets were selected every thirds from 103 soil samples, and the rest two thirds were calibration datasets. Partial least square regression (PLSR) models were obtained to compare the prediction ablity of different spectral proprecessing methods. The best PLS factors in predciton models were determined by the lowest root mean square error (RMSE) in cross validations in calibarion models. The results showed that the absorbtance-transfrom plus Sativazky-Gloay smoothing was the best spectral preprocessing method, and it was the basic of later prediction models.The selection of calibration and validation datasets are crucial for the prediction ability of soil properties. Selecting calibration datasets by soil properties rank (Rank) and Kennard-Stone method (KS) are widely used, and recently a new method named Rank-KS starts to apply in the study. Combining the advantages of Rank and KS methods, Rank-KS makes calibration and validation homogeneous in soil attributive and spectral characteristics. Thus, Rank-KS method impooved prediction accuracy in soil fertility factors.Data mining ability varies with diffdrent chemometric models. For gaining a robust and efficient model, we take linear and nonlinear models into consideration. Linear models include principle component regression (PCR) and PLSR, and support vector machine (SVM) is the nonlinear model. Among three chemometric analysis methods, PCR models gained the lowest prediction accuracy, while PLSR and SVM models have the similar prediction. Black box problem, uncertainty and long run timewere the shortcomings of SVM, so PLSR model is more suitable for application and promotion.OM, TN, AN and pH can be successful predicted using VNIR, MIR and VNIR-MIR, but the prediction accuracy of AP and AK is too low to fulfill the quantitive demand. The prediction accuracy using MIR and VNIR-MIR performed better than VNIR on account of the fundamental vibrationin MIR region and overtones and combinations in VNIR region. MIR model was the most robust among all the models, so it’s the best spectral region for the prediction of soil fertility factors.2. Important bands selection and modeling of soil fertility factors in paddy soilAccording to the prder of VIP socres, one to four important calibration bands were selected. Multiple linear regression models were obtained to predict soil OM and pHusing the selected bands. The results demonstrated it was potential to predict soil OM and pH using a few bands qualitively (RPD>1.6). The MLR model based on 1890 and 1560 cm-1 in MIR region achieved well prediction accuracy (RPD=2.67)on soil OM. When predicting soil pH, VNIR region was more suitable than MIR region, and the model established by 575 and 1100 nm in VNIR region can quantitatively predict soil pH (RPD=1.68).