Mapping Soil Organic Matter By Multi-temporal Remote Sensing Images In Cotton Fields In South Xinjiang

Soil organic matter is one of the important indicators of soil fertility.It not only affects soil formation,soil fertility and soil productivity,but also plays an important role in maintaining global climate stability and ensuring human survival.Traditional soil organic matter monitoring relies on manual sampling and laboratory analysis,which is time-consuming and labor-intensive.The research scope is small,and most of them focus on the field scale or regional scale.There are many limitations in characterizing the change of soil organic matter in regions with strong spatial variability.How to efficiently,quickly and non-destructively obtain soil organic matter information is of great significance for agricultural production and soil quality evaluation.With the development of remote sensing technology,the combination of digital soil mapping and remote sensing technology provides convenience for large-scale soil organic matter monitoring.Remote sensing technology has been widely used in soil organic matter monitoring due to its high timeliness and easy access.However,in areas with high vegetation coverage typically represented by agricultural ecosystems,how to monitor changes in soil organic matter with high precision is still a challenge for remote sensing technology.This study takes the Alar reclamation area in southern Xinjiang as the research area,and takes cotton field soil organic matter as the research object.It is based on the data of the maximum or average value synthesis of the vegetation index in a single period and multi-temporal phases,and the data extracted from the long-term series of Landsat images.Vegetation Index carried out research on soil organic matter in areas with high vegetation coverage.In particular,in the inversion process,this study uses multi-source data collaboratively for the study of soil organic matter.The multi-source data specifically includes soil texture data,proximal sensor data,etc.Modeling methods such as partial least squares（PLSR）,random forest（RF）,and convolutional neural network（CNN）were used to study the change of soil organic matter in the cotton field area,and the soil organic matter content of the cotton field in the study area and the soil organic carbon at a depth of 1m Reserves are mapped.Below are key research findings:（1）Evaluation of SOM in cotton field based on multi-temporal remote sensing dataFirst,soil organic matter was studied based on the vegetation index extracted from single-period and multi-temporal remote sensing images.Among the 48 vegetation indices and their biomass index combinations calculated from satellite imagery,26 vegetation indices and their biomass index combinations had significant correlations with soil organic matter,indicating that all 26 vegetation indices and their biomass index combinations could be used as constructs.Model factor is effectively used in soil organic matter inversion;three machine learning methods show that the model performance of CNN(R²=0.62,RMSE=1.82 g kg^-1,RPD=1.67)is better than that of RF(R²=0.51,RMSE=2.02 g kg^-1,RPD=1.39)and PLSR(R²=0.39,RMSE=2.41 g kg^-1,RPD=1.17);in terms of the distribution characteristics of soil organic matter,the high value of organic matter content is mainly distributed on both sides of the Tarim River and decreases to the north and south sides,and the low value of organic matter content is mainly distributed in the study area The south is close to the desert area and the south of the study area is close to the desert;the factor importance results show that the multi-temporal synthetic biomass index（NDVI＿Sum and NDVI＿Mean）is the most important variable to describe the change of soil organic matter.The results of this study demonstrate the advantages of combining machine learning algorithms with remote sensing technology for soil organic matter retrieval.The results can also provide a scientific basis for agricultural soil management and prevention of soil degradation in the study area.（2）Evaluation of SOM in cotton field based on vegetation,soil and human activity informationThis study proposes a new assessment framework for soil organic matter in cotton fields where returning straw is the main source of soil organic matter,that is,comprehensively considers vegetation,soil and human activity information,compared with the current remote sensing monitoring of farmland soil organic matter that only uses a single vegetation information in terms of accuracy,there will be a significant improvement.Based on the Landsat time series images from 1990 to 2019,the Annual Maximum Biomass Accumulation Index（AMBAI）was developed,and the cultivated land planting years information was quantified for the first time through the spectral index threshold,and then integrated into the near-end sensing Data（soil hyperspectral,soil apparent conductivity）and soil clay content and other information,using partial least squares,random forest,convolutional neural network to establish a prediction model of soil organic matter.The results showed that AMBAI significantly improved the correlation between vegetation information and soil organic matter compared with the single-period biomass index,and the correlation coefficient（R）increased from 0.21 to 0.76.AMBAI,soil hyperspectral,and planting years are the most important factors for evaluating soil organic matter,and their contribution rates are 66.99%,10.72%,and 7.76%,respectively.Relatively speaking,the contribution rates of sand grains and silt grains are low;the contribution rates of vegetation information,soil information and human activity information to soil organic matter are 66.99%,25.25%and 7.76%,respectively.The accuracy of the soil organic matter prediction model constructed by using multiple information sources is significantly higher than that of the single data source model,and the performance of the soil organic matter CNN model integrated with vegetation,soil and human activity information is the best,and its coefficient of determination（R²）,relative analysis error（RPD）,root mean square error（RMSE）are 0.83,2.38 and 1.38 g kg^-1,respectively.This study confirmed that in arid and semi-arid oasis regions,AMBAI and planting years are the most important factors to characterize farmland soil organic matter;the inversion accuracy of soil organic matter can be significantly improved after integrating vegetation,soil and human activity information.（3）Estimation of soil carbon content in cotton field profile based on multi-source data combinationThe combination of the framework based on"vegetation+soil+human activities"information and the deep learning CNN model can be effectively used for the estimation of soil organic carbon storage.The exact coefficient of determination（R²）,relative analysis error（RPD）,and root mean square error（RMSE）are 0.68,0.71g kg^-1,1.60,respectively.The results of factor importance showed that the annual maximum biomass index was the most important variable in evaluating soil organic carbon storage,followed by comprehensive soil information and human activity information,whose contribution rates were 44%,43%,and 13%,respectively.The distribution map of soil organic carbon storage shows that the organic carbon content of1m soil layer in cotton field is about 16.09～65.95 t hm^-2.The distribution characteristics of organic carbon were highly similar to the annual maximum biomass accumulation index and planting years,which indicated that the crop planting process was a process that increased carbon content in the soil environment.This study can provide a scientific basis for subsequent soil management and soil quality assessment.