ObjectiveTraditional digital soil mapping methods are unable to produce detailed soil maps within a reasonable cost and time. Digital soil mapping is a powerful technique, which is popular and widely used by scholars coupled with environmental covariates to map soil types or properties. The selection of environmental covariates is the key to ensuring the accuracy of mapping. Previous studies have proven that remote-sensing images can be used as auxiliary factors for reasoning mapping. Remote sensing data can provide rich soil landscape information, which is consistent with the core idea of using grids to express spatial changes of soil features in digital soil mapping. Moreover, remote sensing technology can obtain real-time information quickly. However, there are few relevant studies on how principal components and texture information of remote sensing factors contribute to the reasoning process. Thus, determining the weight of remote sensing factors in the reasoning process is the key content of this study, which is tested by the reliability of testing mapping results.MethodChengmagang Town, Macheng City, Hubei Province was selected as the study area. Using Chinese soil classification and soil type map with a spacing of 10 meters, which were extracted from the contour data and remote sensing image using a variety of feature selection algorithms to effective screening of variables, this study conducted the soil digital mapping by reasoning machine learning algorithms. Specifically, the recursive feature elimination screening algorithm, ReliefF algorithm and tree-based feature screening algorithm were used to rank all environmental factors in the whole area, plain and hilly areas of the study area, respectively. Then, it screened the effective environmental variables of environmental factors and analyzed the weight of remote sensing factors in the reasoning process. The factors involved in plant-hill region mapping were explored and their importance was determined. According to the selected relatively stable indicators, the gradient boosting decision tree model after parameter tuning of the Bayesian optimization algorithm based on TPE was used for modeling. Also, the mapping accuracy results after different feature screening algorithms were compared between the whole region and the terrain region to further explore ways to improve the accuracy of soil type mapping.ResultThe soil type inference map was verified by 141 independent field sampling sites. The results showed that the importance of remote sensing factors in the plain area was higher than that in the hilly area and the NDVI and Mean values of the remote sensing factors were relatively stable. The highest accuracy of topographical inference mapping based on the recursive feature algorithm was 75.89%, which was higher than the 13.48% and 4.97% of the ReliefF algorithm and tree-based feature screening algorithm, respectively. In addition, among the mapping results of the three feature screening algorithms, the accuracy of the mapping based on terrain factors was higher than that of the overall region mapping. It suggests that remote sensing factors as an auxiliary means to participate in the reasoning process can effectively improve mapping accuracy.ConclusionThis study uses a feature selection algorithm to select features with a strong correlation with soil types as auxiliary variables in the machine learning model. The method is efficient and cost-effective for soil type prediction. Compared to other methods, the soil type mapping method based on machine learning is advantageous and the feature mining and machine learning algorithms have theoretical significance and practical value.