University of Tabriz Journal of Geography and Planning 2008-8078 Articles in Press 2026 02 18 Optimization and Evaluation of a Convolutional Neural Network (CNN) for Gully Erosion Susceptibility Prediction in Sheshtamad county. Optimization and Evaluation of a Convolutional Neural Network (CNN) for Gully Erosion Susceptibility Prediction in Sheshtamad county. 21250 10.22034/gp.2026.69822.3472 FA Nadiya ‌Baghaei Nejad Hakim Sabzevari university Leila Goli Mokhtari Hakim Sabzevari University Abolghasem Amirahmadi Hakim university Ali Beheshti Hakim University Journal Article 2025 10 23 Objective: The aim of this study is to evaluate and optimize a Convolutional Neural Network (CNN) model for predicting gully erosion susceptibility in the Sheshtamad region, using a feature selection approach and hyperparameter optimization.<br /> Methodology: In this research, a total of 600 points were identified through field surveys and satellite imagery, comprising 300 gully and 300 non-gully locations. Of these, 70% of the data were used for model training and 30% for validation. Twenty-one environmental factors including topographic, geological, hydrological, and land-use characteristics were selected as predictive variables. Subsequently, the Mutual Information (MI) criterion and the Random Forest algorithm were applied to identify the most influential factors contributing to gully erosion occurrence.<br /> Next, the CNN model was developed, and hyperparameter optimization was performed to enhance model performance. The optimized model output was then used to generate a gully erosion susceptibility map, which was evaluated in terms of both accuracy and predictive performance.<br /> Given that the study area (Sheshtamad County, western Razavi Khorasan Province) is highly prone to gully erosion, conducting such susceptibility assessments prior to any infrastructural development is essential.<br /> Results: The variable importance analysis revealed that rainfall, distance from rivers, distance from faults, and elevation are the most critical factors influencing gully erosion development. The optimized CNN model, after hyperparameter tuning using the Optuna algorithm, achieved a high prediction accuracy (Accuracy = 0.9338). The resulting susceptibility map indicated that the eastern part of the study area possesses the highest potential for gully erosion occurrence.<br /> Conclusion: The findings demonstrate that the integration of deep learning techniques, particularly the CNN model, with optimal feature selection and hyperparameter tuning provides a powerful framework for accurately predicting gully erosion-prone areas. This approach can support effective soil and water conservation planning and guide sustainable land management policies. Objective: The aim of this study is to evaluate and optimize a Convolutional Neural Network (CNN) model for predicting gully erosion susceptibility in the Sheshtamad region, using a feature selection approach and hyperparameter optimization.<br /> Methodology: In this research, a total of 600 points were identified through field surveys and satellite imagery, comprising 300 gully and 300 non-gully locations. Of these, 70% of the data were used for model training and 30% for validation. Twenty-one environmental factors including topographic, geological, hydrological, and land-use characteristics were selected as predictive variables. Subsequently, the Mutual Information (MI) criterion and the Random Forest algorithm were applied to identify the most influential factors contributing to gully erosion occurrence.<br /> Next, the CNN model was developed, and hyperparameter optimization was performed to enhance model performance. The optimized model output was then used to generate a gully erosion susceptibility map, which was evaluated in terms of both accuracy and predictive performance.<br /> Given that the study area (Sheshtamad County, western Razavi Khorasan Province) is highly prone to gully erosion, conducting such susceptibility assessments prior to any infrastructural development is essential.<br /> Results: The variable importance analysis revealed that rainfall, distance from rivers, distance from faults, and elevation are the most critical factors influencing gully erosion development. The optimized CNN model, after hyperparameter tuning using the Optuna algorithm, achieved a high prediction accuracy (Accuracy = 0.9338). The resulting susceptibility map indicated that the eastern part of the study area possesses the highest potential for gully erosion occurrence.<br /> Conclusion: The findings demonstrate that the integration of deep learning techniques, particularly the CNN model, with optimal feature selection and hyperparameter tuning provides a powerful framework for accurately predicting gully erosion-prone areas. This approach can support effective soil and water conservation planning and guide sustainable land management policies.