AI-Driven Approach Enhances Gully Erosion Prediction and Interpretation

Novel AI methodology improves gully erosion prediction and interpretation

Gully erosion is the most severe form of soil erosion, and it can seriously impact agricultural fields, contributing to sediment loss and severe nutrient runoff into waterways. Gullies can be triggered suddenly by a single heavy rainfall event, creating deep channels that are difficult to rehabilitate even with heavy machinery. Accurately predicting where gully erosion is likely to occur allows agricultural producers and land managers to target their conservation efforts more effectively. In a new study, University of Illinois Urbana-Champaign researchers use a new AI-driven approach that combines machine learning with an interpretability tool to enhance the prediction of gully formation and understanding of these models. They tested the methodology on land in Jefferson County, Illinois. The research is published in the Journal of Environmental Management. "We had conducted a previous study in the same area, but we applied only an individual machine learning model to predict gully erosion susceptibility. While that study provided a baseline understanding, it had limited predictive accuracy. Furthermore, we were not able to explain how the model made predictions. This research aims to address these two key limitations," said lead author Jeongho Han, who recently graduated with a doctoral degree from the Department of Agricultural and Biological Engineering (ABE), part of the College of Agricultural, Consumer and Environmental Sciences and The Grainger College of Engineering at Illinois. Jefferson County is part of the Big Muddy River watershed feeding into Rend Lake. This region features rolling topography and is about 60% agricultural land, primarily used for growing corn and soybeans. The researchers prepared gully erosion inventory maps of the study area based on elevation differences from 2012 and 2015. They also identified 25 environmental variables that can affect erosion susceptibility, including topography, soil properties, vegetation features, and precipitation patterns. Complex environmental processes, such as terrain, soil, hydrology, and atmospheric forces, cause gully erosion, and they are challenging to predict and manage. Machine learning models are increasingly used in erosion prediction, but their accuracy can vary significantly. Stacking multiple models together can improve performance, but adding more models is not enough; it matters how they are combined. The research team evaluated 44 stacked models that combined different features from single models. Next, they created gully erosion susceptibility maps using the best-performing stacking model and four individual models. They found that the best stacking model achieved a prediction accuracy of 91.6%, compared to 86% for the best individual model. To enhance model transparency, the team employed an explainable artificial intelligence (AI) technique called SHapley Additive exPlanations (SHAP). This tool clarifies how different variables influence a model's output, offering deeper insight into AI systems' decision-making process. "When you use AI modeling, you get an output, but it's like a black box. You don't know how it was determined, so you don't have any criteria to evaluate the results. Explainable AI provides metrics that allow you to understand how different variables influence model predictions and how they interact with one another," said corresponding author Jorge Guzman, research assistant professor in ABE. "We integrated a stacking model with SHAP and applied it to a specific land area to demonstrate how it would work. The stacking model improved prediction accuracy, and SHAP helped to interpret what happened within the AI models." For example, the SHAP analysis identified the annual leaf area index of crops as the most influential feature in all base models. Greater leaf coverage reduces the direct impact of rainfall on soil, which in turn decreases the severity of erosion. The proposed framework enables agricultural producers and land managers to interpret AI-model outputs. They can use this information to decide which areas should be managed first and what management practices should be implemented to mitigate soil erosion. "By offering a transparent mechanism to evaluate how different features and models contribute to final decisions, this approach can be extended to broader environmental management and policy-making contexts, facilitating more informed and responsible resource allocation," the researchers conclude in the paper.

Illinois Study: Novel AI Methodology Improves Gully Erosion Prediction and Interpretation | Newswise

Newswise -- URBANA, Ill. - Gully erosion is the most severe form of soil erosion, and it can seriously impact agricultural fields, contributing to sediment loss and nutrient runoff into waterways. Gullies can be triggered suddenly by a single heavy rainfall event, creating deep channels that are difficult to rehabilitate even with heavy machinery. Accurately predicting where gully erosion is likely to occur allows agricultural producers and land managers to target their conservation efforts more effectively. In a new study, University of Illinois Urbana-Champaign researchers use a new AI-driven approach that combines machine learning with an interpretability tool to enhance the prediction of gully formation and understanding of these models. They tested the methodology on land in Jefferson County, Illinois. "We had conducted a previous study in the same area, but we applied only an individual machine learning model to predict gully erosion susceptibility. While that study provided a baseline understanding, it had limited predictive accuracy. Furthermore, we were not able to explain how the model made predictions. This research aims to address these two key limitations," said lead author Jeongho Han, who recently graduated with a doctoral degree from the Department of Agricultural and Biological Engineering (ABE), part of the College of Agricultural, Consumer and Environmental Sciences and The Grainger College of Engineering at Illinois. Jefferson County is part of the Big Muddy River watershed feeding into Rend Lake. This region features rolling topography and is about 60% agricultural land, primarily used for growing corn and soybeans. The researchers prepared gully erosion inventory maps of the study area based on elevation differences from 2012 and 2015. They also identified 25 environmental variables that can affect erosion susceptibility, including topography, soil properties, vegetation features, and precipitation patterns. Complex environmental processes, such as terrain, soil, hydrology, and atmospheric forces, cause gully erosion, and they are challenging to predict and manage. Machine learning models are increasingly used in erosion prediction, but their accuracy can vary significantly. Stacking multiple models together can improve performance, but adding more models is not enough; it matters how they are combined. The research team evaluated 44 stacked models that combined different features from single models. Next, they created gully erosion susceptibility maps using the best-performing stacking model and four individual models. They found that the best stacking model achieved a prediction accuracy of 91.6%, compared to 86% for the best individual model. To enhance model transparency, the team employed an explainable artificial intelligence (AI) technique called SHapley Additive exPlanations (SHAP). This tool clarifies how different variables influence a model's output, offering deeper insight into AI systems' decision-making process. "When you use AI modeling, you get an output, but it's like a black box. You don't know how it was determined, so you don't have any criteria to evaluate the results. Explainable AI provides metrics that allow you to understand how different variables influence model predictions and how they interact with one another," said corresponding author Jorge Guzman, research assistant professor in ABE. "We integrated a stacking model with SHAP and applied it to a specific land area to demonstrate how it would work. The stacking model improved prediction accuracy, and SHAP helped to interpret what happened within the AI models." For example, the SHAP analysis identified the annual leaf area index of crops as the most influential feature in all base models. Greater leaf coverage reduces the direct impact of rainfall on soil, which in turn decreases the severity of erosion. The proposed framework enables agricultural producers and land managers to interpret AI-model outputs. They can use this information to decide which areas should be managed first and what management practices should be implemented to mitigate soil erosion. "By offering a transparent mechanism to evaluate how different features and models contribute to final decisions, this approach can be extended to broader environmental management and policy-making contexts, facilitating more informed and responsible resource allocation," the researchers conclude in the paper. The paper, "Prediction of gully erosion susceptibility through the lens of the SHapley Additive exPlanations (SHAP) method using a stacking ensemble model," is published in the Journal of Environmental Management [DOI: 10.1016/j.jenvman.2025.125478]. Authors are Jeongho Han, Jorge Guzman, and Maria Chu. This research was funded by the US Department of Agriculture through the National Institute for Food and Agriculture (NIFA) award number 2019-67019-29884.