Soil is not merely a passive growing medium. It is a complex physical, chemical and biological system in which water moves through pores, heat travels between layers, gases are exchanged and microbes break down organic matter. A field can shift from water-stressed to saturated within hours after irrigation or heavy rain, while temperature and oxygen availability can alter root activity and microbial respiration. Traditional soil surveys and laboratory measurements remain essential for classification and accurate characterization, but they usually provide snapshots collected at particular locations and times. Digital Pedon is intended to complement—not replace—those measurements by maintaining an evolving computational representation of a soil profile, known technically as a pedon. The software can store relatively stable attributes such as horizon boundaries and soil classification while continuously updating dynamic variables such as moisture, temperature, electrical conductivity and other sensor-derived observations.

At the heart of the framework is a three-layer architecture. The Structural Layer stores the soil’s comparatively stable characteristics, including the geometry of horizons, hydraulic parameters and threshold settings. A soil horizon is a recognizable layer within a profile, such as an organic-rich topsoil or a denser clay-rich subsoil. The Dynamic Layer records incoming measurements in three ways: a key-value store provides rapid access to the current state, an append-only time-series log preserves the profile’s history, and a threshold-event system can trigger alerts when conditions become agronomically important. The Functional Layer contains a registry of mathematical solvers, described by the authors as a “Model Zoo.” Each solver processes the state produced by earlier solvers in a sequence, creating a directed acyclic graph of calculations. Because a solver can only use information already written during the current update cycle, the system prevents circular dependencies without requiring a separate cycle-detection algorithm.

These calculations are designed to translate raw observations into quantities that are much more useful for decision-making. A moisture sensor might report volumetric water content, but growers are often more interested in whether roots can access that water, whether drainage is occurring or whether the soil is approaching a plant’s wilting point. Digital Pedon includes a Van Genuchten–Mualem model to estimate how water is retained and how hydraulic conductivity changes as soil dries. The Van Genuchten equation describes the relationship between soil-water content and matric potential, the pressure state that determines how tightly water is held in soil pores. The associated conductivity formulation estimates how easily water can move through the partly saturated pore network. A Darcy–Buckingham solver then calculates vertical water flux, while a simplified De Vries model estimates thermal properties and a Q10-based solver estimates how temperature and moisture influence respiration. Together, these “software sensors” can infer high-order variables that are difficult or expensive to measure directly.

The framework also tackles a less visible but fundamental problem: soil databases often speak different technical languages. SoilGrids, the U.S. Soil Survey Geographic Database, the European Soil Data Centre and other systems may use different names, units, depth intervals and classification conventions for the same property. Digital Pedon’s Ontology and Normalisation Layer maps incoming terms to machine-readable identifiers associated with the Global Soil Information System and OGC SoilML standards. The package contains a canonical vocabulary covering 41 properties, including 26 core soil properties and 15 sensor-specific variables. Its normalization tools can convert aliases such as “CLAY,” “bdod,” “phh2o,” “SOC_mean” and “VWC” into standardized keys for clay content, bulk density, soil pH, organic carbon and volumetric water content. The system also includes 14 unit conversions, ranging from Fahrenheit to Celsius and microsiemens per centimeter to decisiemens per meter, as well as conversions for some satellite data products.

That common language allows the framework to combine measurements that normally remain isolated. Users can initialize a digital pedon from field observations, laboratory results or existing soil-profile databases. When local information is unavailable, the software can use GPS coordinates to retrieve preliminary profiles from the SoilGrids version 2.0 REST API or the U.S. SSURGO database. It then estimates hydraulic parameters using Rawls–Brakensiek pedotransfer functions—statistical relationships that infer difficult-to-measure properties from more accessible information such as texture and organic matter. In demonstrations at four contrasting locations, the system generated profiles for loam in Ghent, Belgium; sand in Niger’s Sahel; clay in Egypt’s Nile Delta; and silty loam in Italy’s Po Valley. The estimated saturated water content increased from sand to clay, while saturated hydraulic conductivity fell sharply, matching expected physical behavior. The clay profile had an estimated conductivity of 0.8 centimeters per day, compared with 48.2 centimeters per day for the sandy site.

The authors emphasize that these coordinate-based profiles are starting points, not substitutes for site-specific sampling. SoilGrids estimates include uncertainty intervals, and pedotransfer functions can perform differently depending on texture and organic-matter content. In the current software, the median database estimates are treated deterministically, meaning that the model uses a single value rather than propagating the full uncertainty range through every calculation. This is one reason the reported experiments are demonstrations of functionality rather than proof that the framework can make accurate predictions under field conditions. The authors say that quantitative benchmarking against independent observations, device-specific deployments, operational datasets and uncertainty propagation will be priorities for future studies. That distinction is crucial: a digital twin can be technically elegant while still requiring extensive calibration and validation before it should control irrigation or guide high-stakes land-management decisions.

The proof-of-concept experiments nevertheless show how the system could operate. In one synthetic drying-and-wetting cycle, a loamy topsoil was supplied with water contents ranging from 0.08 to 0.43 cubic centimeters per cubic centimeter. The solver chain automatically calculated matric potential, unsaturated hydraulic conductivity and respiration-related outputs at each update. Matric potential changed systematically with moisture, conductivity followed the characteristic nonlinear response expected in partially saturated soil, and modeled carbon dioxide flux reached a maximum near a water content of about 0.26 before declining under wetter conditions. That decline reflects a basic ecological constraint: too much water can fill air-filled pores, restrict oxygen diffusion and suppress aerobic respiration even when microbes are not short of water. In another test, the ontology system correctly resolved all 28 alias-mapping checks across SSURGO, SoilGrids, the European Soil Data Centre and other conventions, with generated identifiers linked to the GLOSIS namespace.

Digital Pedon is also designed to be extended rather than locked to a fixed collection of models. In one example, the researchers added a seasonal heat-conduction solver without changing the framework’s core code. The custom model used a Crank–Nicolson numerical scheme, an implicit finite-difference method commonly used to solve diffusion equations, to calculate temperature through a four-horizon profile extending to 200 centimeters. Thermal diffusivity was supplied dynamically by the built-in De Vries solver at each horizon and timestep, allowing the custom calculation to inherit differences between layers. The simulated profile reproduced expected depth- and season-dependent temperature patterns, including changes in heat transmission at horizon boundaries. The software can run on Python versions 3.9 through 3.12 without compilation or runtime dependencies, making it suitable in principle for small edge devices such as Raspberry Pi, Arduino and ESP32 systems as well as cloud-based geospatial platforms.

The framework includes a threshold-event system that could connect modeled conditions to practical responses. Instead of reacting only to a raw moisture reading, an application might trigger an irrigation or drainage action when the inferred root-zone state indicates saturation, a wilting threshold or excessive salinity. Satellite products such as vegetation indices, land-surface temperature and radar-derived moisture indicators can enter through the same ingestion interface as in-situ measurements, although the researchers caution that syntactic compatibility does not make the observations physically identical. Satellite signals represent areas and often shallow or surface conditions, whereas a probe measures a point at a particular depth. Digital Pedon also exposes its state and solvers through OpenAI-compatible tool schemas, allowing language models to query structured outputs—for example, asking whether an upper soil horizon is under water stress. Such conversational access could make complex models easier to use, but the authors present it as a future decision-support foundation, not as an already validated autonomous agronomic adviser.

By releasing the framework under a Creative Commons Attribution 4.0 license, the researchers hope it will become a shared foundation for precision agriculture, digital soil mapping and environmental monitoring. Potential applications include tracking crop water stress, linking satellite signals to subsurface processes, monitoring soil carbon and assessing land degradation. The longer-term goal is to move beyond isolated digital pedons toward landscape-scale twins that incorporate elevation, slope, curvature, hydrological indices, parent material, soil spectroscopy, hyperspectral imagery and networks of connected sensors. If that vision succeeds, soil data could become both more dynamic and more interoperable: a rainfall event recorded by a field sensor could update a model, alter an estimate of water movement and generate a standardized data product usable across platforms. For now, Digital Pedon is best understood as an open computational substrate—a promising bridge between observations and process-based models whose real-world value will depend on the field tests still to come.

Gully erosion differs significantly from other erosion types because of its sudden onset and severe impact. It typically manifests after intense rainfall events, rapidly creating large channels that disrupt the uniformity of farmland. These gullies not only cause immediate soil loss but also promote sediment runoff, which carries nutrients into adjacent waterways, deteriorating water quality and threatening aquatic ecosystems. The complexity of environmental interactions leading to gully formation has long challenged researchers, especially when trying to foresee which specific land areas will be affected. Traditional prediction models lacked accuracy and explanatory power, leaving farmers and land managers with limited tools to target their conservation efforts effectively.

To address these challenges, the Illinois research team embarked on a study integrating advanced machine learning techniques with innovative interpretability tools. Their approach centers around a stacking ensemble model—a sophisticated AI method that combines multiple machine learning algorithms to boost predictive accuracy. This ensemble approach acknowledges that no single model captures the intricacies of gully erosion on its own, but when carefully combined, they provide a far more precise forecast of erosion-prone zones. The model was rigorously tested within Jefferson County, a predominantly agricultural region characterized by rolling hills and significant corn and soybean production.

The researchers meticulously prepared gully erosion inventory maps by analyzing elevation changes between 2012 and 2015, allowing a temporal lens on where gullies emerged. They then incorporated 25 different environmental variables into their model, encompassing topographical features such as slope and curvature, soil characteristics including texture and organic matter, vegetation indices, and precipitation metrics. This rich dataset was essential for capturing the multifactorial processes driving gully erosion, as terrain, soil, hydrology, and atmospheric conditions interact in complex and non-linear ways.

One of the key insights emerged from comparing the performance of single machine learning models against the stacking ensemble. The best individual model achieved a respectable prediction accuracy of 86%, yet when multiple models were intelligently stacked, the accuracy rose dramatically to 91.6%. This significant improvement underscores the power of ensemble learning frameworks in environmental modeling, where systems are inherently complex and variables interact in nuanced manners. It also highlights that the way models are combined is as significant as the number of models used.

Beyond raw predictive capability, the interpretability of AI models remains a fundamental concern, especially in environmental applications where decision-making benefits from transparency. The Illinois team employed an explainable AI method known as SHapley Additive exPlanations (SHAP). This approach deconstructs model predictions, attributing contributions to individual variables and revealing how they collectively influence outcomes. Applying SHAP allowed the researchers to peer inside the “black box” of AI, identifying which features most substantially impacted the likelihood of gully formation.

Their findings revealed the annual leaf area index of crops as the most dominant variable affecting erosion susceptibility. This metric quantifies the leaf coverage of crop plants and is critical because dense foliage shelters soil from the direct force of raindrops, thereby reducing the detachment and displacement of soil particles. Such biological insights not only validate the model’s predictions but also provide actionable knowledge to land managers aiming to mitigate erosion through targeted crop management and vegetation practices.

The integration of stacking ensemble modeling with explainable AI constitutes a novel framework that marries predictive strength with interpretative clarity. It empowers agricultural stakeholders with a powerful tool that not only identifies high-risk erosion zones but also elucidates the underlying environmental drivers. This fusion enhances trust in AI recommendations by providing rationale that can guide practical conservation decisions, such as prioritizing intervention areas and selecting appropriate soil stabilization strategies.

Jefferson County’s landscape, with its variability in topography and extensive agricultural use, served as an ideal testbed for this approach. The success here suggests broader applicability in diverse environmental contexts where gully erosion threatens soil health and water quality. By offering a transparent and accurate prediction system, this methodology has the potential to transform soil conservation efforts on regional and national scales.

The research also signals a pivotal moment for environmental modeling by demonstrating that machine learning does not need to remain an opaque technology. Instead, through tools like SHAP, AI can become a collaborative partner in environmental science, illuminating complex interactions and enhancing our capacity to manage natural resources responsibly. These advances are poised to influence policy-making by providing scientific evidence that officials can rely upon for allocating resources and designing sustainable land use plans.

Funded by the U.S. Department of Agriculture’s National Institute for Food and Agriculture, this study bridges cutting-edge AI science with on-the-ground agricultural challenges. Its outcomes pave the way for smarter, more precise environmental stewardship that aligns with modern technology’s promise. As climate change and land use pressures intensify, such predictive and explainable tools will be indispensable for ensuring the longevity of productive soils and the health of the ecosystems they support.

In conclusion, the University of Illinois team has forged a new pathway in environmental modeling by coupling stacking ensemble machine learning techniques with explainable AI methods. Their work not only elevates the precision of gully erosion susceptibility predictions but also demystifies the AI decision-making process, enabling targeted conservation efforts and fostering sustainable agricultural management. This research stands as a testament to the potential of AI to tackle complex environmental problems with both power and transparency, charting a hopeful course for soil preservation amidst dynamic natural and human systems.

Subject of Research: Prediction of gully erosion susceptibility using AI-driven stacking ensemble models and explainability techniques

Article Title: Prediction of gully erosion susceptibility through the lens of the SHapley Additive exPlanations (SHAP) method using a stacking ensemble model

News Publication Date: 25-Apr-2025

Web References:
https://doi.org/10.1016/j.jenvman.2025.125478

References:
Han, J., Guzman, J., & Chu, M. (2025). Prediction of gully erosion susceptibility through the lens of the SHapley Additive exPlanations (SHAP) method using a stacking ensemble model. Journal of Environmental Management. https://doi.org/10.1016/j.jenvman.2025.125478

Image Credits: Marianne Stein, University of Illinois

Keywords: Agriculture, Environmental sciences, Modeling, Soil erosion, Machine learning, Explainable AI, Gully erosion, Stacking ensemble, SHAP