Gully erosion represents a severe environmental challenge where concentrated water flow carves deep channels into landscapes, stripping away fertile soil at alarming rates. This phenomenon devastates agricultural productivity, silts waterways, and exacerbates climate change through carbon release.
Predicting where gullies will form has historically been difficult due to complex interactions between rainfall patterns, terrain steepness, soil composition, and vegetation cover. Traditional physical models require intensive field measurements that often prove impractical across large areas.
Earlier machine learning approaches offered partial solutions but suffered from inconsistent accuracy and opaque decision-making. A groundbreaking 2025 study published in the Journal of Environmental Management (Volume 383) has transformed this field by combining advanced AI with crystal-clear explanations, achieving unprecedented prediction accuracy while revealing exactly how environmental factors interact to cause erosion.
Developing the Gully Erosion Predictive Framework
The research team from the University of Illinois pioneered a two-tiered approach centered on Jefferson County, Illinois—a 1,512 km² agricultural region experiencing significant erosion pressure. First, they meticulously mapped actual gully locations using high-resolution LiDAR elevation data from 2012 and 2015.
By analyzing elevation drops exceeding 50 cm (accounting for agricultural tillage depth) and cross-referencing with aerial imagery, they identified 1,000 verified gully sites and 1,000 non-erosion control points.
This rigorous methodology ensured highly accurate training data for their models. Next, the team evaluated 25 environmental variables spanning topography, soil properties, vegetation cover, and precipitation patterns.
Through statistical analysis, they eliminated redundant factors like seasonal vegetation indices that duplicated annual measurements.
The remaining ten most influential predictors were identified using mutual information scoring, which measures how much each variable reveals about erosion risk.
Annual precipitation emerged as the strongest predictor with a score of 0.199, followed by soil organic matter (0.074), profile curvature (0.060), and annual leaf area index (0.045). Slope angle, though critical, ranked tenth with a score of 0.042, highlighting how multiple factors combine to trigger erosion.
The Power of Stacking Ensemble Models
To overcome the limitations of single-algorithm models, the researchers developed 44 variations of a “stacking ensemble” framework. This sophisticated approach uses multiple machine learning models in two layers.
Four base models—Random Forest (RF), Gradient Boosting Machine (GBM), Logistic Regression (LR), and Deep Neural Networks (DNN)—independently analyzed the environmental data.
Their predictions were then fed into a meta-learner model (also chosen from RF, GBM, LR, or DNN), which synthesized the base models’ insights into a final, refined prediction.
Performance testing revealed dramatic differences among approaches. The best standalone model, Gradient Boosting, achieved an Area Under the Curve (AUC) score of 0.86—already impressive for complex environmental modeling.
However, the optimal stacking ensemble (using RF and GBM as base models with LR as the meta-learner) achieved a remarkable AUC of 0.916. This 6.6% improvement translates to significantly more reliable risk mapping.
For perspective, AUC measures prediction accuracy from 0.5 (random guessing) to 1.0 (perfect precision), making 0.916 exceptional in real-world applications. Furthermore, this ensemble model demonstrated superior precision (0.91) and recall (0.90), minimizing both false alarms and missed high-risk zones.
Mapping Risk with Precision And Gully Erosion Predictions with SHAP
Applying this optimized model across Jefferson County yielded transformative insights. The resulting susceptibility map classified agricultural land into four risk categories. Crucially, 33% of farmland—totaling 89,208 hectares—was identified as “very high risk,” requiring immediate conservation measures.
Another 6% fell into “high risk,” while 7% was “moderate risk.” The remaining 54% was deemed “low risk,” allowing farmers to focus resources efficiently.
Comparisons with standalone models highlighted the stacking ensemble’s superiority. The Logistic Regression base model overestimated threats, classifying 87% of farmland as “very high risk”—an impractical outcome for conservation planning.
The Deep Neural Network model showed similar overestimation at 55%, while the generally reliable Gradient Boosting model identified only 27% as critical zones, potentially overlooking vulnerable areas. Only the stacking ensemble balanced accuracy with practicality, producing a nuanced risk map that aligns with ground observations.
The study’s second breakthrough came through SHAP (SHapley Additive exPlanations), an explainable AI technique that demystifies how models reach conclusions. SHAP quantifies each environmental factor’s contribution to predictions, both overall and for specific locations.
Globally, it revealed that annual leaf area index (LAIann)—a measure of vegetation density—was the most influential factor. Higher LAIann values consistently reduced erosion susceptibility, confirming vegetation’s role in shielding soil.
However, this effect weakened significantly on steeper slopes, explaining why some vegetated areas remain vulnerable. SHAP also uncovered complex interactions invisible to traditional models.
For instance, profile curvature (measuring land concavity/convexity) showed a U-shaped relationship with erosion.
Flat areas had low risk, but susceptibility spiked at convexity values of +0.3 and -0.3. Furthermore, when combined with high Stream Power Index (indicating water flow force), risk intensified dramatically on convex slopes.
Perhaps most valuably, SHAP clarified how the stacking ensemble resolved disagreements between base models. In cases where RF predicted high risk but GBM did not, the meta-learner prioritized GBM’s input due to its proven accuracy, preventing overestimation.
Practical Applications and Global Implications
Farmers in high-risk zones can now implement targeted interventions. Planting winter cover crops like rye or clover boosts LAIann, directly countering the dominant risk factor. On convex slopes exceeding 4° gradient, contour plowing or terracing disrupts water concentration.
The 89,208 hectares identified as critical could be prioritized for USDA conservation programs, maximizing the impact of limited resources.
Policymakers benefit from SHAP’s transparent explanations, which justify conservation spending with clear cause-and-effect relationships. Statements like “Cover crops reduce risk here because LAIann overrides moderate slope steepness” carry more weight than opaque algorithmic outputs.
The framework’s adaptability extends beyond gullies. Flood prediction could integrate river discharge data, landslide models could add seismic metrics, and water quality monitoring could incorporate fertilizer application records—all while maintaining explainability.
Conclusion
This research marks a paradigm shift in erosion prediction. By fusing stacking ensemble models with SHAP explainability, the team achieved both unprecedented accuracy (AUC 0.916) and actionable transparency. Farmers gain precise maps distinguishing critical zones from lower-risk land, while SHAP’s insights guide effective interventions.
As climate change intensifies rainfall patterns, such tools become indispensable for protecting soils and food security. The methodology’s scalability offers hope for erosion hotspots worldwide, proving that AI can illuminate solutions to our planet’s most pressing environmental challenges—one explainable prediction at a time.
Key Terms and Concepts
What is Gully Erosion: Gully erosion is the process where running water removes soil along natural drainage lines, creating deep channels or gullies in the land. These gullies are larger and deeper than small rills and cannot be smoothed out by normal farming activities. Gully erosion happens when water runoff concentrates in small channels and scours the soil, often after heavy rains or storms. It is important because it causes loss of fertile topsoil, damages farmland, and increases sediment in rivers and lakes, which harms water quality. For example, in agricultural areas where vegetation is removed, gullies can form quickly and spread, making the land unusable for farming.
What is Runoff: Runoff is water, usually from rain, that flows over the land surface instead of soaking into the soil. When soil is bare or compacted, more runoff occurs, which can concentrate in drainage lines and cause erosion. Runoff is important because it carries soil particles away, leading to erosion and sedimentation in water bodies. For instance, after a heavy storm, runoff can create or deepen gullies by washing away soil.
What is Leaf Area Index (LAI): Leaf Area Index is a measure of the total leaf area of plants per unit ground area. It indicates how much vegetation covers the land surface. LAI is important because dense vegetation protects soil from erosion by intercepting rainfall and slowing runoff. Higher LAI values usually mean lower erosion risk. For example, a forest with a high LAI will have less soil erosion compared to bare farmland.
What is Slope: Slope refers to how steep the land surface is, usually measured in degrees or percent. It affects how fast water flows over the land. Steeper slopes cause faster runoff, which increases the potential for erosion, including gully erosion. Slope is a key factor in erosion models. For example, gullies are more likely to form on steep hillsides than on flat land.
What is Stream Power Index (SPI): Stream Power Index is a topographic measure that estimates the erosive power of flowing water at a specific location. It combines the slope of the land and the amount of water flow (drainage area). Higher SPI means more energy to erode soil. SPI is used to identify areas prone to erosion and gully formation. For example, channels with high SPI values are more likely to develop gullies.
What is Topographic Wetness Index (TWI): Topographic Wetness Index measures how likely soil is to be saturated with water, based on slope and the upstream contributing area. Areas with high TWI tend to be wetter and may have slower runoff, which can affect erosion patterns. TWI helps predict where water accumulates and where gullies might form or grow.
What is Soil Organic Matter: Soil organic matter is the decomposed plant and animal material found in soil. It improves soil structure, fertility, and water retention. Organic matter is important because soils rich in organic matter are less prone to erosion. For example, soils with low organic matter are more easily washed away by runoff, increasing gully erosion risk.
What is Hydraulic Conductivity: Hydraulic conductivity is a measure of how easily water moves through soil pores. Soils with high hydraulic conductivity allow water to soak in quickly, reducing runoff and erosion. Conversely, low hydraulic conductivity leads to more surface runoff and higher erosion risk. For example, clay soils often have low hydraulic conductivity and are more vulnerable to erosion.
What is Bulk Density: Bulk density is the mass of soil per unit volume, including the spaces between soil particles. High bulk density means compacted soil with fewer pores, which reduces water infiltration and increases runoff. Compacted soils with high bulk density are more prone to erosion. For example, soils compacted by heavy machinery or livestock grazing often have high bulk density and higher erosion risk.
What is Clay Content: Clay content refers to the percentage of clay particles in soil. Clay soils hold water well but can become easily compacted and may disperse when wet, making them vulnerable to gully erosion. Soils with high clay content may develop tunnel erosion that leads to gullies. For example, dispersive clay soils in some regions are hotspots for gully formation.
What is Precipitation (Rainfall): Precipitation is any form of water, like rain or snow, falling to the ground. The amount, intensity, and timing of rainfall greatly influence erosion. Heavy, intense storms cause more runoff and erosion than light, steady rain. For example, in Illinois, bimodal rainfall peaks in May and November contribute to gully erosion.
What is Maximum Daily Precipitation: Maximum daily precipitation refers to the highest amount of rainfall recorded in a single day over a period. It is important because extreme rainfall events can trigger sudden gully formation by causing rapid runoff and soil detachment. For example, a storm with 50 millimeters of rain in one day can cause more erosion than several days of light rain.
What are Ephemeral Gullies: Ephemeral gullies are small channels formed by concentrated runoff that can be easily filled by tillage but tend to reform after subsequent rain events. They are important because they represent early stages of gully development and contribute significantly to sediment transport. For example, ephemeral gullies often form along farm tracks or field edges and indicate areas needing conservation.
What are Permanent Gullies: Permanent gullies are deep channels too large to be removed by normal farming operations. They cause long-term loss of land and require costly repairs. These gullies develop from ephemeral gullies if left unchecked. For example, a gully one meter deep and several meters wide is considered permanent and severely impacts farm productivity.
What is Headward Erosion: Headward erosion is the process where a gully or channel lengthens upslope, cutting back into the land. This causes gullies to grow larger over time. It is important because it increases the area affected by erosion and can threaten infrastructure. For example, a gully starting at a small depression can extend upslope by headward erosion during repeated storms.
What is Rill Erosion: Rill erosion is the removal of soil by small, shallow channels less than 30 centimeters deep. Rills are the first stage of concentrated flow erosion and can develop into gullies if not managed. Rill erosion is important because it indicates early soil loss and can be controlled by tillage or vegetation. For example, after a rainstorm, small rills may appear on bare slopes.
What is Soil Texture: Soil texture describes the proportion of sand, silt, and clay particles in soil. It affects water retention, infiltration, and susceptibility to erosion. Sandy soils drain quickly but may be loose, while clay soils hold water but can be compacted. Soil texture helps determine erosion risk and management practices. For example, loamy soils with balanced texture are generally less prone to erosion.
What is Vegetation Cover: Vegetation cover refers to the plants growing on the soil surface. It protects soil by intercepting rainfall, reducing runoff speed, and stabilizing soil with roots. Good vegetation cover is the best natural defense against erosion. For example, grass cover on a slope greatly reduces the chance of gully erosion.
What is Sediment Yield: Sediment yield is the amount of soil material transported from a catchment area by runoff into rivers or reservoirs. It is important because high sediment yield indicates severe erosion and can reduce water quality and reservoir capacity. For example, gullies contribute significantly to sediment yield in agricultural watersheds.
What is Drainage Area: Drainage area is the land area that contributes runoff to a specific point in a drainage network. Larger drainage areas collect more water, increasing the potential for erosion downstream. Drainage area is used in calculating stream power and erosion risk. For example, a gully at the bottom of a large drainage area may have more erosive power.
What is Bimodal Rainfall: Bimodal rainfall means there are two distinct periods of high rainfall in a year. This pattern affects erosion because it creates two peak runoff seasons, increasing the chances of gully formation and expansion. For example, Jefferson County, Illinois, experiences bimodal rainfall peaks in May and November, influencing erosion patterns.
What is Machine Learning (ML): Machine learning is a type of artificial intelligence where (early) computers learn patterns from data to make predictions or decisions without being explicitly programmed. ML is important in erosion studies because it can model complex, nonlinear relationships between environmental factors and erosion risk. For example, ML models can predict where gullies are likely to form based on terrain, soil, and rainfall data.
What is Stacking Ensemble Model: A stacking ensemble model combines multiple machine learning models by training a meta-model to integrate their predictions. This approach improves accuracy by leveraging the strengths of different models. In erosion prediction, stacking helps produce more reliable susceptibility maps. For example, combining Random Forest and Gradient Boosting models with Logistic Regression as a meta-learner increased prediction accuracy.
What is Random Forest (RF): Random Forest is a machine learning algorithm that builds many decision trees and combines their results to improve prediction accuracy and reduce overfitting. RF is widely used in environmental modeling because it handles complex data well. For example, RF can classify land areas into erosion risk categories based on multiple input variables.
What is Gradient Boosting Machine (GBM): Gradient Boosting Machine is a machine learning technique that builds models sequentially, where each new model corrects errors of the previous ones. It is powerful for prediction tasks with complex data. GBM was one of the best-performing models for gully erosion prediction in the study.
What is Logistic Regression (LR): Logistic Regression is a statistical model used for binary classification problems, estimating the probability of an event occurring. It is simple, interpretable, and often used as a meta-learner in stacking models. For example, LR combined predictions from RF and GBM to make final erosion susceptibility decisions.
What is SHapley Additive exPlanations (SHAP): SHapley Additive exPlanations is an explainable AI method that assigns importance values to each input feature in a machine learning model. It helps users understand how each variable affects the prediction, increasing transparency and trust. For example, SHAP showed that leaf area index and slope were the most influential factors in predicting gully erosion susceptibility.
Reference:
Han, J., Guzman, J. A., & Chu, M. L. Prediction of Gully Erosion Susceptibility Through the Lens of the Shapley Additive Explanations (Shap) Method Using a Stacking Ensemble Model. Available at SSRN 4854201. https://doi.org/10.1016/j.jenvman.2025.125478
