Research On Remote Sensing Estimation Model Of Soil Organic Matter

Soil organic matter（SOM）is an important part of soil,which is closely related to the formation history,structure and ecological function of soil.As one of the global carbon pools,SOM plays a regulatory role in atmospheric carbon dioxide concentration.Meanwhile,SOM provides the basic nutrients needed by vegetation and affects vegetation productivity.Therefore,SOM is closely related to global climate change and human food issues,and rapid dynamic monitoring of SOM has become the key technical issue of the above issues.Traditional SOM monitoring is a chemical analysis method based on soil samples from limited sampling points.How to correlate SOM monitoring results of limited sampling points with soil spectral information,soil landscape elements information and remote sensing images to deduce SOM status and SOM dynamics in sampling area is a challenge in the field of soil remote sensing.Based on the laboratory SOM content and hyperspectral reflectance information of soil samples from Haitan Island in 2013,SOM characteristic spectra were selected by using a variety of spectral processing methods（Savitzky Golay smoothing filter S-G,multiple scattering correction MSC,continuum removal CR,first-order differentiation after three transformations and spectral index after resampling）.Geographically weighted regression（GWR）,Support Vector Machine（SVM）and random forest（RF）were used to estimate SOM,and the best hyperspectral estimation model of SOM was obtained.Based on the theory of soil genesis,based on the data of Landsat OLI,DEM,land use and soil type,16 environmental factors which have a close impact on SOM were extracted.The variables were selected by geographical detector model,and the SOM content was inversed by GWR,SVM and RF models to obtain the optimal regional SOM estimation and fine mapping.The spatial distribution map of soil SOM in 2020 was further predicted,and the regional SOM dynamic monitoring was obtained.The results show that:（1）After S-G smoothing filtering,MSC,CR and first-order differential processing,MSC transform can best improve the correlation between reflectance and SOM（R²=0.921,RMSE=0.252,RPD=3.552 of tranining dataset;R²=0.519,RMSE=0.685,RPD=1.553 of validation dataset in RF modle）.Compared with the hyperspectral optimal transform MSC,the spectral index constructed by wide band after indoor hyperspectral resampling has better prediction effect on SOM content（the training dataset R²=0.940,RMSE=0.219,RPD=4.073;the verification dataset R²=0.597,RMSE=0.627,RPD=1.644under the optimal RF model）.（2）Considering the influence of vegetation and soil moisture,ND₁₂ has a good correlation with SOM content（the correlation coefficient is0.526）,which can be used as SOM index to improve the prediction ability.（3）Compared with the traditional Kriging interpolation method,the remote sensing environmental factor method can reflect the spatial distribution of regional SOM accurately and reasonably after adding environmental factor covariates.In addition to the first level,the number of patches in the second to sixth levels increased at least 117 times.（4）Among the three models based on point domain and region,RF model has the strongest SOM prediction ability.Regional RF model training dataset R²=0.731,RMSE=0.496,RPD=1.926;validation dataset R²=0.532,RMSE=0.887,RPD=1.456.（5）Compared with2013,the content of SOM in 2020 shows an overall improvement trend,and the improvement proportion is significantly larger than the degradation proportion（the maximum conversion is from the second level to the third level,and the conversion area is 42.70 km²）.The improvement and degradation are closely related to the change of land use types.The technical methods obtained in this study can improve the spatial prediction accuracy of SOM,realize the rapid dynamic monitoring of SOM,and provide basic data for regional soil carbon estimation,regional agricultural yield estimation and global climate change research.