Enhancing machine learning-based gully erosion susceptibility modeling using a prediction-oriented spatial cross-validation framework

• Spatial CV (kNNDM) improves generalizability of Machine Learning-based modeling • LULC is the dominant gully driver, with bare and cultivated lands at highest risk • SVM outperforms alternative models in gully erosion susceptibility prediction • Integrating LULC with robust susceptibility maps supports targeted watershed planning Prediction where gullies are likely to form is essential for sustainable land management in fragile ecosystem. Machine learning (ML) has advanced gully erosion susceptibility modeling by capturing complex, non-linear geospatial relationships. However, most ML applications often rely on random cross-validation, a non-spatial approach that inflates model performance metrics and limits predictive generalization. This study aims to optimize ML-based gully erosion susceptibility modeling in Upper Awash River Basin using a spatially explicit k-fold Nearest Neighbor Distance Matching (kNNDM) cross-validation framework. Five ML algorithms: Random Forest, Gradient Boosting Machine, Naïve Bayes (NB), Artificial Neural Network (ANN), and Support Vector Machine (SVM) were trained using 584 georeferenced gully and non-gully points derived from field surveys and high-resolution imagery. Both spatial and random partitioning tested with training (80 %) and validation (20 %) sets. Feature selection identified land use/land cover as the most influential gully-conditioning factor. Models trained with spatial cross-validation produced more realistic and unbiased performance estimates compared to inflated metrics from random partitioning. Among all algorithms, SVM achieved the best balance of predictive accuracy and generalizability (Accuracy = 0.788, ROC-AUC = 0.831, F1 = 0.783), followed by ANN (Accuracy = 0.735, ROC-AUC = 0.844 and F1 score = 0.747), while NB performed low due to its simplistic assumptions of feature independence in such complex geomorphological processes. The resulting susceptibility map delineates gully-prone zones ranging from low to very high risk, within Areas of Applicability. Cultivated and bare landscapes exhibited the highest susceptibility, highlighting the need for targeted, spatially informed intervention. The suitability of kNNDM framework demonstrates a reliable spatial validation approach for gully erosion modeling in data-scarce regions, enhancing predictive reliability and supporting evidence-based land management and conservation planning.