Construction Of The Hyperspectral Prediction Model Of Soil Organic Matter

With the continuous improvement of hyperspectral measurement technology and study of multivariate methods using in specific areas, hyperspectral technology has applied more and more widely. Application in agriculture study constantly enriching and deepening, obtained many good results. Compared with conventional chemical method, using hyperspectral techniques to predict soil major nutrients, could avoid time-consuming, laborious, costly, environmental pollution and other issues, achieve large-scale, fast, non-destructive monitoring. With the improvement of modeling methods, models which could achieve simultaneous monitoring of a variety of physical and chemical properties continue to be made, so that the application of hyperspectral technology advantages in terms of agricultural soils.Predictive ability and stability were the main symbol of evaluating the quality of hyperspectral prediction model. Because hyperspectral technology has features of high resolution, multi-band, large amount of data, so that technology can be used to express fine information, however it bring more demand while prediction models. Selection of pre-processing methods and calibration techniques directly affect the accuracy of modeling. Based on this, selected25pre-processing transformations, combined with two major modeling methods to predict Henan four main soil types:fluvo-aquic soil, cinnamonic soil, lime concretion black soil and paddy soil. With soil modeling comparison, according to the predictive ability and stability of the model, select the best predictive model of different soil types. The most appropriate pre-processing transformations for each modeling method and soil types were summarized. The results show that sifted soil for indoor hyperspectral measurement, the use of derivative method and a variety of pre-processing transformation in combination, can improve the modeling accuracy. Based on different spectral data different data smoothing points have different effects on modeling accuracy. And not all the data pre-processing method can improve the modeling accuracy, by trying contrast could obtain the best prediction model. The organic sensitive band which selected by different modeling methods are different, but the main focus bands are similar, The results showed that fluvo-aquic soils organic matter sensitive bands were focused on the VIS wavelength band400-600nm, and NIR band800nm, and three water absorption bands1400nm,1900nm,2200nm nearby. Cinnamonic soil organic matter sensitive bands were focused on the VIS620nm-700nm, and nearby NIR920nm,1820nm water absorption bands2200nm. Lime concretion black soil organic matter sensitive bands were focused on the VIS500-600nm, near by780nm, and NIR1320nm,1800nm, and water absorption band2250nm and2400nm. Paddy soil organic matter sensitive bands are mainly focused on VIS500-700nm, nearby NIR1700nm,1800nm,2400nm and water absorption band2250nm. Moisture absorption of visible light and near band2250nm is the four co-sensitive band of soil organic matter.In order to establish a more practical prediction model, undisturbed fluvo-aquic soil from Long-term experiment was specifically selected with different gradients soil organic matter content and different soil property. After flooding-low temperature drying operation, to obtain simulated field soil moisture and different nutrient status samples for measurement and modeling of indoor and outdoor spectrum. The model predictive ability and stability are high, especially in the coefficient of determination of soil moisture prediction models, test models Rv2reached0.955, the RMSEv was only0.026(n=356), the model can be used to predict the fluvo-aquic soil organic matter content and moisture of different soil fertility status and soil moisture conditions. Different pre-processing methods used in models were compared and selected. The results show that compared to prediction models use of sieved soil, the optimal pre-processing method for spectral modeling use undisturbed soil were different. Pre-processing method which could eliminate system noise and spectrum difference caused by sample surface, such as multiple scatter correction, standard normal transformation, established the most accurate prediction models. However the derivative method which could get a high prediction models using indoor sieved soil, especially second derivative method, was not suitable for undisturbed soil spectral model, use of derivative method enhanced spectral noise, had the risk of reduce model accuracy. Meanwhile, in this paper356indoor and253outdoor undisturbed samples which has different levels of soil moisture and organic matter were divided into3levels according to soil organic content and5levels according to soil moisture, then established their adaptation3soil moisture prediction models and5organic prediction models. These models had higher accuracy, and the model selected into smaller bands, the prediction model is simple, which make it get required spectral data more easier, so that increased the usefulness of the model, these models can be used to predict the fluvo-aquic undisturbed soil organic matter and soil moisture.Analysis of the broadly representative spectrum of undisturbed soil which collected from indoor and outdoor conditions summed up the difference of sensitive hyperspectral bands which collected from two conditions. The result shows that sensitive hyperspectral bands selected by different models has similarities, but the bands selected to field model were less, VIS500-700nm, nearby NIR900nm were undisturbed fluvo-aquic soil organic sensitive band, VIS400-580nm bands were undisturbed fluvo-aquic soil moisture common sensitive band.