Soil Mapping In Mixed Areas Based On Feature Selecting And Machine Learning Methods

In digital soil mapping,the accuracy of obtaining soil-environment knowledge directly determines the final precision of soil mapping.In flat areas,traditional environmental factors such as topography and vegetation have small variations,making it difficult to effectively reflect the spatial variability of soil.Extracting environmental covariates that can effectively reflect the spatial variation of soil in flat areas becomes a key and challenging research focus in digital soil mapping,as well as an important scientific issue for improving the accuracy of soil mapping in flat areas.Based on the aforementioned issues,this study comprehensively utilized topographic factors and remote sensing images as sources of knowledge for inference mapping.It focused on the mixed region of hills and plains to study digital soil prediction mapping methods and analyzed the importance of remote sensing factors in the inference process.The development of this research can expand the theoretical and methodological knowledge acquisition in digital soil mapping,lay a foundation for improving mapping accuracy in flat areas,and have significant research value and significance.This study focused on Chengmagang Town in Macheng City,Hubei Province,which was a typical mixed region of plains and hills.Based on topographic and remote sensing factors as the fundamental data,different feature selection methods were employed to analyze the importance of environmental factors.The study investigated the impact of different environmental factors on the spatial distribution of soil types in the plains and hills,and determined the optimal set of environmental factors.Finally,various machine learning methods were used for modeling and inference mapping.The objective of this research was to comprehensively utilize topographic and remote sensing factors as sources of knowledge for inference mapping,thereby improving the accuracy of soil mapping in flat areas.The specific research contents were as follows:(1)In this study,different feature selection algorithms,including Recursive Feature Elimination(RFE),Relief F,and Tree-based feature selection algorithms,were used to select environmental factors.The results showed that the environmental factors selected based on the RFE algorithm achieved higher prediction mapping accuracy compared to those selected using the Relief F algorithm and the Tree-based feature selection algorithm.Moreover,the various selection algorithms allowed for ranking the importance of the environmental factors.The generated importance ranking results indicated that Normalized Difference Vegetation Index(NDVI)and Mean had greater importance in the plains region compared to the overall region and the hilly region.This further suggested the high importance of remote sensing factors in soil mapping in the plains region.(2)The environmental factors selected through different feature selection algorithms were input into various machine learning models to predict soil types and ultimately obtain an inference map of soil types.Furthermore,the application of different machine learning models in mapping the overall region,plains region,and hilly region was explored.The results showed that the soil inference maps generated by different models exhibited finer granularity compared to traditional soil maps.In comparison to mapping the overall region,mapping based on terrain subdivision achieved higher accuracy and better results.(3)In this study,the spatial distribution of soil types in the research area showed significant similarities in the prediction results of these three models.However,the accuracy and stability of the three models varied.From a model perspective,the Ada Boost classification model,GBDT classification model,and Bagging classification model each demonstrated their advantages in classification mapping.In terms of mapping accuracy,the GBDT classification model exhibited the highest accuracy when generating soil maps based on terrain subdivision.From the perspective of model stability,the Ada Boost classification model showed the most stable accuracy when generating soil maps based on terrain subdivision.In terms of overall model performance,the Bagging classification model performed well in generating soil maps based on terrain subdivision while ensuring stability requirements.In conclusion,combining machine learning methods with feature selection algorithms for soil mapping modeling,prediction,and inference holds significant research value.