<![CDATA[Landslide susceptibility modeling is a critical task for disaster mitigation, yet it is frequently undermined by a severe class imbalance inherent in landslide datasets, where non-landslide instances vastly outnumber actual landslide events. This imbalance leads to biased machine learning models with poor predictive power for the minority (landslide) class, resulting in unreliable hazard maps. This study, focusing on the high-risk area of Malang Regency, Indonesia, addresses this challenge by proposing an innovative framework that integrates a Generative Adversarial Network (GAN) for synthetic data augmentation with a Radial Basis Function Network (RBFN) for classification. A highly imbalanced dataset with a 1:10 ratio of landslide to non-landslide points was constructed to establish a realistic baseline. On this data, the RBFN model, while theoretically powerful for capturing non-linear relationships, failed completely, achieving a Recall of 0.00 for the landslide class. The novelty of this research lies in the specific application of a GAN, trained for 15,000 epochs, to generate high-fidelity synthetic landslide data, thereby creating a perfectly balanced training set. After retraining on this augmented data and undergoing a systematic hyperparameter tuning process, the RBFN’s performance was dramatically transformed. The optimized model achieved an F1-Score of 0.9333 and a Recall of 0.8750, elevating its performance from total failure to a level competitive with the robust Random Forest benchmark. This work validates that the integrated GAN-RBFN approach is a highly effective methodology for overcoming the data imbalance problem in geospatial hazard modeling. By turning a previously unreliable classifier into a powerful predictive tool, this method has significant practical implications for developing more accurate landslide susceptibility maps, which are crucial for informed spatial planning and enhancing early warning systems.]]>
Copyrights © 2026
