High-Precision Inversion Method For Soil Organic Matter In Cultivated Land Based On Erosion Zoning

Soil organic matter,which contains large amounts of carbon and nitrogen,is a key attribute of soil quality.It plays an important role in soil health,structure,fertility,and plant growth.Soil erosion is one of the primary causes of soil degradation and loss of soil organic matter,significantly impacting the redistribution process of surface soil organic matter.Currently,most studies on soil organic matter loss due to soil erosion focus on micro-scale investigations,such as fields and slopes.These studies examine the impact of various topographic conditions,cropping practices,rainfall intensity,and other factors on the process of soil organic matter loss at a smaller scale within the study area.Such studies are often influenced by the regional scale and are challenging to apply to various regions,particularly in large-scale study areas with intricate topography and geomorphology.Furthermore,most studies on regional-scale soil erosion assessment rely on empirical models.The assessment results are influenced by the method of parameter calculation,data accuracy,regional characteristics,and other factors.This complexity makes it challenging to uncover the process of regional-scale soil erosion on soil organic matter loss.In order to address these issues,this study focused on Northeast China as the research area.It utilized a deep learning model to simulate the spatial and temporal dynamics of soil erosion and investigated how soil erosion impacts the accuracy of the soil organic matter inversion model on a large scale.Considering that soil organic matter inversion models constructed in areas with varying erosion levels may differ significantly,a partitioning modeling strategy was adopted in the soil organic matter modeling process.Combining the advantages of random forest,geographically weighted regression,and convolutional neural network,various soil organic matter inversion models were developed in regions with varying levels of erosion.The spatio-temporal differentiation characteristics of soil organic matter and its influencing factors were also analyzed.The main research results of the paper are as follows:（1）The spatial and temporal dynamics of soil erosion were simulated by integrating multi-source remote sensing data with the Transformer model.The dynamic changes in soil erosion in Northeast China over the past 40 years were evaluated using the Revised Universal Soil Loss Equation to provide data support for soil erosion simulation.The mapping relationship between environmental variables and soil erosion was established by combining multi-source data and the Transformer model.A soil erosion simulation model was constructed using a data-driven approach to simulate historical and future changes in soil erosion.The results show:soil erosion is relatively serious in the Northeast China,and since 1999,there has been a trend of mitigation in soil erosion.The soil erosion simulation obtained using short-term climate change data（3-year average temperature and average precipitation）was more accurate compared to long-term climate change.Slope,temperature,and precipitation are important factors that contribute to soil erosion,particularly in regions with slopes greater than 6°in the Northeast China.There is a significant variation in soil erosion trends under different future climate scenarios.Dramatic climate change exacerbates the risk of soil erosion,especially under conditions of high social vulnerability and high radiative forcing from anthropogenic interventions.（2）Combining the advantages of random forest,geographically weighted regression,and convolutional neural network models,a few-shot convolutional neural network soil organic matter inversion model was constructed using the Stacking integration method.This approach effectively enhanced the inversion accuracy of the few-shot neural network model.Based on the Stacking integration theory,the random forest and geographically weighted regression models were chosen as the base models.The convolutional neural network was employed as the meta-learning model to construct the soil organic matter inversion model using satellite remote sensing data and environmental variables.The spatial and temporal differentiation characteristics of soil organic matter in Northeast China were analyzed,along with the driving factors of soil organic matter changes.The study also discussed the dynamic changes of soil organic matter under various land use types and topographic landscapes.The results show that the model combines the unique advantages of nonlinear fitting of random forest,local linear fitting of geographically weighted regression,and automatic learning of convolutional neural network.This combination effectively improves the accuracy of the small-sample deep learning soil organic matter inversion model.Compared with the results of soil organic matter inversion using random forest,geographically weighted regression,and convolutional neural network alone,the accuracy of the soil organic matter model constructed based on the Stacking integration theory was significantly enhanced.The coefficient of determination（R²）increased from 0.561 to0.642,and the root mean square error（RMSE）decreased from 13.72 g kg^-1 to 12.39 g kg^-1.The results of the geodetector and random forest importance analysis showed that temperature,brightness,and the visible-light band have significant effects on the accuracy of soil organic matter inversion.Soil organic matter in Northeast China has shown an overall decreasing trend,with noticeable variations in the decline among different land use types.The rate of soil organic matter depletion was more pronounced from 1985 to 2000,but it slowed down from 2001 to 2020.For soil organic matter in different landform types,organic matter content is higher in tablelands and lower in hills.（3）According to the soil erosion simulation model,various erosion zones were identified.Utilizing a partitioning modeling approach,a few-shot convolutional neural network model was employed to predict the soil organic matter content of cultivated land and analyze its spatial and temporal variations across different erosion zones.According to the erosion modulus,the Northeast region was divided into two erosion zones:erosion zone 1 and erosion zone 2.A soil erosion simulation model was used to determine the zoning boundaries and probabilities of different erosion zones.The soil samples were divided into two groups based on different erosion zones.Different few-shot convolutional neural network models were constructed to invert soil organic matter for each erosion zone using the partitioning modeling method.Finally,the results of soil organic matter inversion from different erosion zones were combined using a hybrid probability model.The results show:the combination of a hybrid probability model and Stacking integration theory can significantly enhance the accuracy of few-shot learning models.The R² value increased from 0.582 to 0.715,while the RMSE decreased from 13.39 g kg^-1 to 10.83 g kg^-1.The use of partitioning modeling and hybrid probability models can effectively improve the accuracy of soil organic matter inversion models.The partitioning modeling method considers the variations in the impacts of different erosion zones on the accuracy of soil organic matter estimation.The results of the geodetector showed temperature,elevation,erosion modulus,L-factor,and band reflectance had a significant impact on the results of soil organic matter inversion.From the perspectives of global sensitivity and model interpretability,the variables influencing the model accuracy varied significantly across different erosion zones.In erosion zone 1,temperature,precipitation,and elevation had a more pronounced impact on the model accuracy.Conversely,in erosion zone 2,the ratio index exerted a greater influence on the model compared to the difference index and the normalized difference index.