Predicting Of Forest Soil Organic Carbon Content Based On Near Infrared Spectroscopy

Forest soil organic carbon (SOC) plays an important role in the soil fertility that indicates nutritional status of timbered soil. Rapid prediction of SOC content will become a significant research that in real time estimates the dynamic change of forest soil carbon reserves. With the advantages of rapid, easy, and nondestructive, near-infrared spectroscopy (MRS) coupled with partial least squares (PLS), principal components regression (PCR), back propagation (BP) neural network and support vector regression (SVR) was investigated in this paper to predict SOC content. Application of various pretreatment methods, optimization of spectrum, principal component extractions and model parameter were applied for model optimization.(1) Different pretreatment methods and optimization of spectrum were used for model optimization. The effect of model was compared PLS with PCR algorithm. Results showed that predictive effect was the best, when spectral region was1380-1450nm,1800-1950nm,2050-2300nm, pretreatment methods were Savitzky-Golay (SG), multiplicative scatter correction (MSC) and first-order derivative (FOD), principal component was8and applying PLS algorithm. Carbon content was predicted with the RMSE, SEP and determination coefficient of0.5143,0.5140and0.7537for the calibration set. Those meet the accuracy requirement.(2) Eight typical principal components were extracted from principal component analysis (PCA) with the application of establishing prediction model of BP. Results showed that Levengerg-Marquardt (LM) algorithm was applied to achieve desirable modeling performance. For the test set, the correlation coefficient (R) was0.8829and the root mean square error (RMSE) was0.4120, for the validation set, the R was0.7800and the RMSE was0.5002, for the train set, the R was0.8494and the RMSE was0.4538. Those meet the accuracy requirement and have improved.(3) To rapid predict forest soil organic carbon content, best parameters c and g and different normalized ways and kernel functions were applied for optimization then SVR was used for implementation of nonlinear fitting. Results showed that when [-1,1] normalization method, RBK function were used and the best parameter c and g were0.5and0.0625, predictive effect with R was0.8903and RMSE was0.2739. Those meet the accuracy requirement and have further improved.(4) PLS, BP and SVR model all could achieve the rapid prediction of forest soil organic carbon content. By comparison, PLS could realize a rough prediction, BP neural network was more accurate, and SVR was the most accurate. This paper provided a new train of thought for the determination of forest soil organic carbon content, which provides the theoretical basis and technical support for effectively raising management.