Inversion Modeling Of Soil Organic Carbon In The Western Lakeside Oasis In Bosten Lake Based On Hyper-Spectral Data,China

Soil organic carbon（SOC）was an important indicator to measure the balance of oasis agriculture and soil ecosystem in arid areas.It affects the formation and stability of soil structure to a large extent,and it was the mitigation and improvement of soil salinization,land Pollution and desertification were important factors in environmental issues.Due to the extensiveness of soil spatial distribution,the diversity of types,and the differences in physical and chemical properties,rapid acquisition of soil information and large-area digital mapping and updating of soil were particularly important.The emergence of hyperspectral remote sensing provides powerful technical means for the estimation of soil organic carbon content and the detection of soil environment,which can quickly obtain soil attribute information in this area with high efficiency and large area,and meet the needs of agricultural production guidance.Taking the lakeside oasis on the west bank of Bosten Lake as the research area,a total of 255 soil samples were collected according to different seasons and different land use types.Based on the measured SOC content and hyperspectral data,the seasonal changes of the SOC content and the changes of different soil depths in the study area Characteristics.Combined with spectroscopic analysis methods to reveal the characteristics of soil spectral reflectance of different SOC content,different land use types and different soil depths.Based on the analysis of SOC content and the original spectrum and the spectral reflectance after pretreatment,the correlation coefficient method,successive projections algorithm（SPA）and genetic algorithm（GA）screen the characteristic band of SOC content,and used partial least squares regression（PLSR）,support vector machine（SVM）and random forest（RF）regression models were constructed to invert the hyperspectral estimation model of SOC content.Three indicators verify the accuracy and stability of the model,and discuss the accuracy and feasibility of vis-NIR spectroscopy combined with regression methods to quickly estimate the SOC content of the oasis.The main conclusions were as follows:（1）The range of 0～50 cm SOC content in the study area was 0.75～78.76 g·kg^-1,and the coefficient of variation was between 61.53～82.83%,showing medium variability,log-normal distribution,and with increasing soil depth.Deepen the SOC content shows a decreasing trend.The SOC content at different soil depths in spring,summer and autumn showed moderate variability,and the coefficient of variation was between 43.97%and 94.94%.Comparing the mean SOC content in three different seasons:summer>spring>autumn.The SOC content of cultivated land was 0.92～40.92 g·kg^-1,and the coefficient of variation was 38.89～70.23%,showing moderate variability.The SOC content of cultivated land in the 0～40 cm soil layer decreases with the increase in soil depth.After the 10～20 cm soil layer,the variability of grassland SOC content gradually weakens,and the variability of unused land SOC content was higher than that of the other three types of land use.The highest ratio was 115.69%.（2）The shape of the SOC spectrum curve for the different soil layers was relatively constant.The spectral reflectance was higher with the increase of soil depth in the same waveband,and the spectral reflectance of the SOC content of grassland in the same waveband was the highest,followed by unused land,and woodland spectrum the reflectance was the lowest.The SOC content and the original spectral reflectance showed a negative correlation,with a correlation coefficient of-0.62<r<-0.07.There were 1160 bands that have passed the extremely significant test（P<0.01）,mainly concentrated in 524～1299 nm,1469～1790 nm and 1973～2056nm bands have the highest correlation in the 661 nm band,with a correlation coefficient|r|of0.62.After preprocessing,the number of bands that pass the extremely significant test（P<0.01）was reduced to among the 414 bands,the correlation coefficients were the highest in the 788 nm,800 nm and 1768 nm bands,with r all greater than 0.80.（3）Comparing the accuracy of the PLSR model of 10 preprocessed spectra,the optimal spectral preprocessing algorithm was Savitzky-Golay smoothing（SG）,standard normal variate（SNV）and first derivative（1st Der）,that was,based on the SG-SNV-1st Der algorithm,the established PLSR model modeling set and verification set R² were 0.75 and 0.80,respectively,and the RPD was 2.01,which had good predictive ability.（4）The number of characteristic bands selected by the original spectral data using the correlation coefficient method,SPA algorithm,GA algorithm and GA-SPA algorithm were 5,19,140,and 4,respectively,accounting for 0.28%,1.06%,7.78%and 0.22%respectively of the number of bands.After spectral preprocessing,the number of characteristic wavelengths selected by using correlation coefficient method,SPA algorithm,GA algorithm and GA-SPA algorithm are 5,14,216 and 9,respectively,which account for 0.28%,0.78%,12.01%and 0.22%.（5）The RF model has the best modeling effect,and its accuracy was higher than that of the SVM model and the PLSR model.After the original spectrum was preprocessed by SG-SNV-1st Der,the accuracy of the model constructed by the five spectral variables of the full band,correlation coefficient method,GA algorithm,SPA algorithm and GA-SPA algorithm was higher than that of the original spectrum,and the RF model R² was improved respectively increased 1.32%,2.50%,1.20%,6.10%and 13.58%.A comprehensive analysis of the accuracy and stability of the five characteristic variable models shows that the model effect was:GA-SPA>SPA>GA>correlation coefficient>full band.The RF model constructed with the characteristic variables selected by the GA-SPA algorithm can better estimate the SOC content of the study area.The R² of the modeling set and the verification set were 0.91,0.92,and the RMSE was 6.25 g·kg^-1,6.18 g·kg^-1,RPD was 2.45.