Research On Hyperspectral Soil Organic Matter Prediction Models Of Songnen Plain

Accurate hyperspectral Soil Organic Matter(SOM) prediction model is the basic requirement of SOM content fast determination, precision agricultural and Soil Organic Carbon estimation.Accurately soil distribution information is basic step of soil imaging and soil database building.Meanwhile, the information will provide guidance of land management and land use. The prediction models of hyperspectral remote sensing of soil are used in SOM, iron oxide and moisture fast determination significantly. Partial Least Squares Regression(PLSR) is widely used in estimation of soil physical and chemical parameters such as Soil Organic Matter(SOM) and moisture content, due to its advantages in dealing with collinearity of variables like hyperspectral reflectance. However, it is hard to determine optimal input variables for a PLSR model of SOM such as, mathematical transformation of spectral reflectance(MT), wavelength ranges and spectral resolution. Laboratory hyperspectral reflectance of soils in Songnen plain were analyzed in this study, and the Orthogonal Experimental Design(OED) method for deriving optimal input variables for SOM prediction models was introduced. Results show that 1) the OED method can be an effective way to identify the optimal input variables of a PLSR model for SOM prediction, and the optimal combination input is 1400-2450 nm with derivative of the logarithmic reciprocal reflectance(DLRR) at 25 nm for Black soil, and 400-2450 nm with DLRR at 5nm for different soils, the optimal MT of both of them is DLRR. This study provides a new approach to determining optimal input of a PLSR model, and enhances the comparability of different research results. 2) 13 Key Points(KP) were extracted from correlation coefficient curve which is calculated between SOM and spectral reflectance with different MT as model input. Results show that KP as PLSR model input is an effective way to reduce the number of inputs, and have similar ability of prediction with OED(RPD=3.12, 2.33). This method provides a new approach for SOM prediction. PLSR has better accuracy with KP as model input compared with MSR model, and both of them can be used in Black SOM fast determination(RPD>3.0). However, model prediction ability is the same when applied in different soils(2.0<RPD<2.5). 3) Soil spectral characteristics may be different due to diverse soil physiochemical properties and parent material, therefore,model accuracy of different soils normally lower than single soil. Fuzzy k-means based on hierarchical analysis(FKMH) divided soils type as five clusters, and the classification is conform to the samples were collected. The model accuracy is improved from 0.76 and 2.04 to 0.83 and2.42 after soil classification. 4) Another method spectral angle mapper based on characteristic spectrum(SAMC) divided spectral reflectance into three types. Results show that both model accuracy and stability are improved to 0.85 and 2.60 with using SAMC, and the model has good ability to predict SOM. To sum up, soil classify first with quantitative parameters instead of putting them into prediction model directly will beneficial to acquire higher SOM prediction accuracy. Meanwhile, SAMC provides a new theory and method to soil classification.