【Objective】Digital soil mapping is a burgeoning and efficient method to express the spatial distribution of soil. Based on a data mining algorithm, this method establishes a soil-landscape relationship model to infer soil mapping by using raster data as an expression and computer-assisted. The key to improving the accuracy of digital soil mapping is constructing a suitable soil-landscape relationship model. However, the commonly used methods of digital soil mapping cannot meet the application requirements of soil mapping given the complicated nature of terrains consisting of plains and hills. How to fully consider the main links of the soil-landscape relationship model to accurately infer the spatial distribution of soil types needs further discussion. 【Method】The northern part of Chengmagang town, Macheng City, Hubei province was selected as the study area. It was divided into two terrain units, plains and hills. Based on the 28 environmental variables, Decision Tree (DT), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost) were used to select optimal mapping methods for each region. Then, the optimal variables combination was selected according to the factor importance ranking of each region. Moreover, the optimal mapping methods were used to establish a soil-landscape relationship model linking soil types to the optimal variable combinations, upon which soil type mapping was inferred for each region. Soil-type mapping results for plain and hilly areas were combined as the soil-type mapping result of the terrain region. Finally, the mapping accuracy of the whole region was compared with the terrain region to further explore ways to improve the accuracy of soil-type mapping in Plain and Hill Mixed Regions. 【Result】Under different terrain conditions, the performance of each inference mapping method was different as well as the optimal inference mapping method. The performance of RF and XGBoost was superior to other algorithms. Specifically, the RF performed better in whole and plain regions while the XGBoost was the best algorithm in the hill region. The model accuracy was further effectively improved through variable screening, with the maximum increase of overall accuracy and Kappa coefficient being 4.96% and 0.059 in the whole region, respectively. However, the model accuracy improvement was not obvious in the plain region, with the increase of overall accuracy and kappa coefficient being 1.43% and 0.018, respectively. Also, the increase in overall accuracy and kappa coefficient was 2.82% and 0.03 in the hill region. Compared with the whole mapping method, the inference mapping method based on terrain zoning had the highest accuracy, and the overall accuracy and Kappa coefficient were 73.05% and 0.69, respectively. Meanwhile, the plain region required more remote sensing factors to participate in inference mapping compared to the whole and hill regions. 【Conclusion】The inference mapping accuracy in plain and hill regions can be effectively improved by optimizing the mapping method, selecting environment variables, and adopting appropriate mapping way. This study can provide some references for the screening of environmental variables, the selection of mapping algorithms, and the construction of mapping ways of inference mapping in plain and hill regions. It provides promising and practical examples and technical support effective for promoting the improvement of the accuracy of inference mapping in complex terrain areas.