Comparison Of Estimation Methods For Obtaining Soil Properties From Hyperspectral Data In The Sanjiang Yuan Regions

149 surface soil samples(0 ~ 30cm)were collected from Sanjiang Yuan regions,and then soil properties(total nitrogen, total carbon, carbon-nitrogen ratio, organic matter, pH, silicon, aluminum, iron, magnesium, manganese, arsenic, lead, zinc, cadmium and chromium) and soil reflectance spectral were respectively analyzed and scanned in the laboratory. After three smoothing methods including Nine Weighted Moving Average(NWA),Savitzky-Golay(SG),Wavelet Transform(WT) and first derivative transformation of the reflectance,characteristics wave bands of major soil properties were extracted by correlation analysis(CA), stepwise multiple linear regression(SMLR) and genetic algorithms(GA) in combination with partial least squares(GA-PLS) method, BP neural network(GA-BPNN)method and support vector machine(GA-SVM) method.Then the hyperspectral estimation models were built,validated and compared based on characteristics wave bands and full wave bands respectively. To determine the best model of each soil properties and to select the best method of spectral smoothing method and the optimum extracted characteristic bands method.The mail results were as follows:(1)The soil spectral reflectance was smoothed using Savitzky-Golay convolution smoothing, wavelet transform and nine points weighted moving average methods. The results showed that the wavelet transform(WT) in this study was more suitable for removing the noise of the spectral reflectance of the soil samples in this study, compared with other two kinds of smoothing methods(Savitzky-Golay convolution smoothing and nine points weighted moving average)(2)The estimation model between the full wave bands spectra and the concentrations of 17 kinds of soil propertites were established based on PLSR, BPNN and SVM methods respectively. It was found that the order of the model estimated capacity is that SVM model > BPNN model > PLSR model. It was not a simple linear relationship between the concentrations of soil propertites and spectrum, so the PLSR model showed some limitations, but BPNN and SVM models can better deal with the nonlinear relationship between spectrum and the concentrations of soil propertites;(3) The three kinds selection methods of feature wave bands including correlation analysis(CA), stepwise multiple liner regression(SMLR) and genetic algorithms(GA), were respectively combined with PLSR, BPNN and SVM models to establish statistic correlation between soil properties concentrations and spectral reflectance for predicting soil properties. The results indicated that the estimation capacity of GA is better than that of SMLR and CA for PLSR; GA is better than that SMLR and CA for BPNN; SMLR is better than that GA and CA for SVM;The extraction methods of feature bands corresponding different modeling methods were not completely same, and the appropriate selection methodsof characteristicwave bandsshould be selected according to the modeling method. But on the whole, genetic algorithm and multiple stepwise regression analysis of these two methodsof selecting characteristic wave bands were the best, and they can replace the traditional correlation analysis method to modeling.(4) Compared with soil properties estimation models from the whole band, the accuracies of estimation models from the feature band are lower as a whole,but it needs less bands with a lower data redundancy,and simple model structure,what’s more, it can run quickly and effectively,and that showing more stability accuracies, so it can replace the whole wave bands model to estimate the soil properties in this study area.(5) Synthesizing the estimating model of whole wave bands and characteristic wave bands, then it was found that all models had excellent ability to estimate the total nitrogen, the total carbon and the organic matter, but do not have the ability about silicon, manganese, copper and mercury. Most of models had a rough ability ofestimating carbon nitrogen ratio, pH value, aluminum, iron, magnesium, arsenic, lead, cadmium and chromium. In this study area hyperspectral remote sensing for most estimation of soil chemical propertites played an effective role, but it was not suitable for estimating silicon, manganese, copper and mercury in the soil.