Field verification
Field verification of accuracy is crucial for assessing the practical value of soil type prediction models. This validation method involves verifying the predicted soil types by collecting ground-truth samples in the field to test the accuracy of the models under real-world conditions. Such validation based on real-world data provides a more accurate reflection of the model’s performance compared to relying solely on theoretical or simulated data.
The accuracy of the calibration soil type map is 79.38%, while the accuracy of the SVM model is only 81.67%, which is slightly higher than the calibration soil type map. This may be due to its limitations in handling high-dimensional data and complex spatial relationships. The XGB model achieved an accuracy of 86.83%, which could be attributed to the uneven sample distribution and the complexity of the field terrain, which limited its predictive capability. In contrast, the RF performed significantly better, with an accuracy of 88.35%, surpassing the calibration soil type map. This improvement may be attributed to RF’s robust handling of large-scale and multivariate spatial data, using multiple decision trees to improve prediction accuracy and stability. The highest accuracy was achieved by the VEM, which reached 93.75%. This significantly higher accuracy compared to other models indicates that the VEM effectively handles the complexity and diversity of soil type prediction by combining the strengths of multiple algorithms.
Results of the consistency analysis
Different types of soil develop a series of unique physical and chemical characteristics during their formation, influenced by their origin, formation conditions, and long-term exposure to factors such as climate, biology, and topography49. Aeolian soil, for instance, is a loose, immature soil with only A-C horizons, formed in arid and semi-arid regions on sandy parent material. It is at the initial stage of soil development, with weak soil-forming processes, consisting mainly of fine sand that can easily be moved by the wind50. The pH of aeolian soil is slightly alkaline, ranging between 8 and 951. As the aeolian soil transitions to the fixed aeolian soil stage, vegetation continues to develop, with coverage increasing to over 30%52. The biological soil-forming processes become more evident, and the soil profile further differentiates, leading to the formation of a thicker crust or humus-stained layer. This stage sees the accumulation of organic matter, giving the soil a certain level of fertility, which can further develop into the corresponding zonal soil53,54. In Tongzhou District, sand meadow fixed aeolian soil exists, as shown in Table 2. The average pH of sand meadow fixed aeolian soil in Tongzhou District is 8.22, meeting the pH consistency requirements.
Cinnamon soil is a type of semi-leached soil that forms under warm temperate, semi-humid monsoon climates and in environments characterized by arid forests and shrub-grass steppe vegetation55. It develops through processes of clayification and calcification, resulting in a clay-enriched B horizon and the accumulation of CaCO3 (pseudo-mycorrhizae) at certain points within the soil profile. The soil tends to be neutral or slightly acidic56. In the Tongzhou District, both aquic cinnamon soil and developed cinnamon soil are present. Aquic cinnamon soil, influenced by groundwater, displays rust spots or iron-manganese nodules in the lower part of the profile, and it is slightly alkaline, with a soil pH greater than 7.0. As shown in Table 2, the pH of aquic cinnamon soil areas is slightly alkaline, which meets consistency requirements. Developed cinnamon soil, found in the warm temperate semi-humid region, is subject to severe erosion, leading to poorly developed soil profiles that are shallow and contain numerous gravel or rock fragments, making biological enrichment challenging. In this study, the RSEI is used to represent the degree of biological enrichment. As indicated in Table 2, the RSEI value for developed cinnamon soil areas in Tongzhou District is only 0.43, the lowest among all subtypes, indicating a low degree of biological enrichment, which meets consistency requirements57.
Fluvo-aquic soil is a semi-hydromorphic soil that forms under the influence of groundwater, characterized by a profile structure that includes a humus layer (cultivated layer), an oxidation-reduction layer, and a parent material layer56. In the Tongzhou District, fluvo-aquic soil, dehumidified soil, wet-flush soil, and salinized fluvo-aquic soil are present. Fluvo-aquic soil, developed on the sloping plains of alluvial plains transitioning from gentle hills to depressions, is generally neutral to slightly alkaline. As shown in Table 2, the pH levels in fluvo-aquic soil, dehumidified soil, and wet-flush soil regions are slightly alkaline, meeting pH consistency requirements.
Dehumidified soil, a type of fluvo-aquic soil with leaching and deposition characteristics, represents a transitional type from fluvo-aquic soil to adjacent non-hydromorphic or semi-hydromorphic soils. It is typically found in relatively elevated terrains such as gentle hills on plains, natural levees of ancient river channels, and high terraces, with a deeper groundwater Table58. Wet-flush soil, characterized by gleying features, is a transitional type from fluvo-aquic soil to bog soil. The average elevation of dehumidified soil is relatively high, while that of wet-flush soil is relatively low, satisfying elevation consistency tests59.
Salinized fluvo-aquic soil is a type of fluvo-aquic soil in which the surface layer’s soluble salt concentration reaches salinization thresholds. It represents a transitional type from fluvo-aquic soil to saline soil and is distributed in relatively low-lying or marginal areas of alluvial plains24. It often forms a mosaic distribution with fluvo-aquic soil and wet-flush soil. In addition to the primary soil-forming processes of fluvo-aquic soil, it also undergoes salinization. As shown in Table 2, the total ion content in salinized fluvo-aquic soil is the highest. The sample data from this region indicate the presence of sulfate-dominated salinization. Specifically, the equivalent ratio of sulfate ions (SO₄²⁻) to chloride ions (Cl⁻) in the soil is 3.82:1.29, exceeding 1, confirming its classification as sulfate-salinized fluvo-aquic soil, in line with consistency tests. Ion data from other sampling points in the Tongzhou District, as shown in Table S2, do not meet the conditions for saline soil.
Causes of soil type evolution
The disappearance of bog soil
In the past, bog soil in Tongzhou developed in low-lying areas with long-term water accumulation and under hydrophilic vegetation. As a result, bog soil was primarily distributed along riverbanks and in local depressions formed by human excavation60. For bog soil to form, the groundwater level must be within 1 m of the surface. The groundwater level in Tongzhou’s alluvial plain was historically 1.5 to 2.0 m, while in the low-lying alluvial plain, it was 0.5 to 1.0 m, and in transitional depressions, it was 0.5 to 1 m—conditions suitable for bog soil formation61. However, following China’s reform and opening-up, urban development in Tongzhou has led to increased groundwater extraction, causing a decline in the water table. Additionally, urbanization has resulted in the expansion of impervious surfaces, disrupting the complete water cycle and further lowering the groundwater level. The increase in impervious surfaces has also altered the local microclimate in Tongzhou, with changes in average precipitation and other conditions making it difficult for organic matter to accumulate, leading to changes in soil types. According to the 2020 groundwater depth data in Tongzhou, the maximum groundwater depth reached 3.1 m, which is no longer conducive to the development of bog soil. Moreover, Tongzhou has been significantly impacted by human activities. These activities have altered the physical and chemical properties of bog soil, leading to its evolution and eventual disappearance.
The disappearance of cinnamon soil
In the western part of Tongzhou, within the remnants of gently sloping hills on the edge of the Quaternary alluvial fan, cinnamon soil developed from loess-like parent material. The higher elevation and good drainage conditions of these areas made them conducive to the formation of cinnamon soil. At the edge of the alluvial fan or in locally flat, low-lying areas, the terrain is gentle with slight slopes, and groundwater influences the lower parts of most cinnamon soils to varying degrees, thus playing a role in the modern soil-forming processes62. Due to Tongzhou’s location on the edge of the alluvial fan, cinnamon soil exhibits transitional characteristics towards fluvo-aquic soil in the plain areas. The local topography and groundwater levels significantly affect the development of cinnamon soil. In the stabilized dune areas on the alluvial plain, where the terrain rises 3–8 m above the surrounding plain, relatively stable soil development conditions exist, making these areas the main but scattered distribution zones for developed cinnamon soil24. However, because the terrain of the alluvial fan edge gradually transitions to the plain, aquic cinnamon soil occupies a larger area within the distribution range of cinnamon soil, while cinnamon soil only appears sporadically in locally elevated areas. Human activities, such as land leveling and excavation, have disrupted the natural soil profile configuration by removing the topsoil and exposing the subsoil, thereby altering the natural environment required for cinnamon soil formation. Additionally, the lowering of the groundwater table has contributed to the transformation of cinnamon soil into other soil types.
The increase in the distribution of Wet-Flush soil
As shown in Fig. 5, in the transition areas between the western edge of the alluvial fan and the low-lying alluvial plain in Tongzhou District, the gradual elevation of the terrain due to the continuous infilling and silting of sediments has led to a relative lowering of the groundwater table and a significant reduction in the extent of seasonal waterlogging. The soil has long moved away from marshy conditions and has been developing towards fluvo-aquic soil, now referred to as wet-flush soil. The figure also indicates that in the plain areas, river flooding, variations in flow velocity, and the differing sedimentation processes caused by water sorting have resulted in multiple changes in river courses and interbedded deposits. These processes have led to various sediments frequently overlapping, creating interlayers of sand and clay. This directly affects the dynamics of soil water and salt content, as well as the degree of soil development, resulting in a mosaic distribution of different soil types. The relatively low terrain at the edges of the transition areas is the main distribution zone for wet-flush soil. Even in slightly elevated areas within the plains, these locations are still considered relatively low-lying in terms of overall terrain trends63.
Elevation and Groundwater Depth with wet-flush soil Distribution.
The decrease in the distribution of salinized Fluvo-Aquic soil
In the past, the southern region of Tongzhou had scattered distributions of salinized fluvo-aquic soil, particularly at the margins of transition areas between the edges of low-lying depressions and the expansive plains. The formation of this soil type was closely related to a higher groundwater table and intense evaporation. In these southern regions, the groundwater level was generally higher than in typical fluvo-aquic soils, and the soil texture tended toward sandy loam. Under poor drainage conditions, this made the soil more prone to salinization64. However, due to effective long-term water resource management, land use planning, land reclamation, and soil improvement measures in Tongzhou District, the accumulation of salts in the surface layers has been significantly reduced. The lowering of the groundwater table has further improved the soil environment, leading to the transformation of salinized soil into other soil types.
The decrease in the distribution of aeolian soil
Aeolian soil was primarily distributed along rivers and ancient riverbanks, where it was transported by wind and accumulated in dune-like formations across the plains. Along the riverbanks, aeolian soil often exhibited a parallel, band-like distribution or accumulated on the inner edges of river meanders65. In some areas along the river, larger expanses of aeolian soil were found. Additionally, aeolian soil often accumulated near villages, forming dune-like structures that rose 2–3 m above the surrounding plains and extended in ridges along the prevailing wind direction. However, following extensive agricultural land development and comprehensive planning and management efforts, much of the aeolian soil has been leveled and converted into basic farmland or forest land. As the terrain in aeolian soil distribution areas gradually flattened, these soils began transitioning into sand fluvo-aquic soil. The reduction in aeolian soil in Tongzhou District reflects this evolution, with the land predominantly transforming into sand calcareous fluvo-aquic soil, consistent with the natural progression of aeolian soil.
Distribution of soil sections
The soil distribution in Tongzhou is significantly influenced by geomorphology, topography, and hydrogeological conditions, leading to a highly varied soil distribution across different regions. The micro-scale distribution patterns of soil types are also quite pronounced. Human activities such as irrigation, drainage, fertilization, and land leveling have further altered the soil properties, making the distribution of soils more complex. Tongzhou is located in a warm temperate semi-humid region on the edge of the Beijing alluvial fan, including a large alluvial plain in the southeast that slopes toward the southeast. The western part, being at the edge of the alluvial fan, has higher terrain and is the primary distribution area for cinnamon soil. In contrast, the southeastern alluvial plain, with relatively flat terrain, is primarily influenced by sediment type, groundwater activity, and human cultivation, making it the main distribution area for fluvo-aquic soil. Due to the sloping terrain, variations in local elevation and hydrological conditions lead to differing soil distribution patterns. As shown in Fig. 6, the extensive overlap and widespread distribution of different soil types across latitudes, longitudes, and elevations suggest that their spatial patterns are influenced by a complex interplay of environmental factors. It is evident that sand calcareous fluvo-aquic soil and loam calcareous fluvo-aquic soil have the broadest distribution, with light loam bottom clay calcareous fluvo-aquic soil being more densely distributed in the central part of Tongzhou. The areas marked 15–18 and 41 represent the distribution of cinnamon soil, all located at relatively higher elevations. The areas marked 1–6 represent the distribution of dehumidified soil, which is found in relatively high and higher elevation areas.
Soil distribution profile (The left chart shows the relationship between soil types and latitude, while the right chart shows the relationship between soil types and longitude. The vertical axis represents elevation. The density of points indicates the frequency of occurrence of each soil type).
Limitations and future work
The results of this study not only demonstrate the potential of ensemble learning in improving soil type prediction accuracy but also emphasize the importance of integrating multiple machine learning algorithms. By leveraging the strengths of different models, this approach exhibits significant advantages in enhancing predictive accuracy, managing uncertainties, and improving robustness, thereby providing more detailed soil information and contributing to a better understanding of soil evolution.
However, this study has some limitations and also provides new directions for future research. Firstly, one of the main limitations of this study is the size of the study area. The current study was conducted in a specific, limited area, which may limit the generalisation ability of our model66. Soil types and environmental conditions may vary significantly from region to region; therefore, the validity of our model in other regions has not been validated67. In order to improve the generalisation ability and applicability of the model, future research should consider testing and validating the model in a wider and diverse geographical area45.
Regarding the direction of future research, considering the potential of machine learning in soil mapping, exploring different ensemble methods would be a valuable endeavour. The current model employs a voting integration strategy, but other ensemble techniques, such as stacking and blending, may provide superior performance40. Stacking methods involve combining the predictions of several different machine learning models, while blending incorporates a meta-learner to optimise the final prediction during the model integration process. Future research could explore these different ensemble techniques to determine which combination best suits the complexity and diversity of soil types68. Additionally, while deep learning models have demonstrated advantages in large-scale soil mapping, their application here was constrained by the limited sample size and the need for interpretability in soil evolution analysis12. Future studies will incorporate CNN-based architectures with expanded datasets to evaluate their comparative performance in dynamic soil type prediction18. This may involve the development of more flexible models that can easily integrate new data sources and environmental indicators to reflect changes in the environment in real time69.
