Rainfall-conditioned landslide susceptibility mapping in data-scarce regions using an interpretable deep convolutional neural network: a case study of the Jebel Marra volcanic massif, Sudan.

Catastrophic landslides in the Jebel Marra volcanic massif of western Sudan recently caused multiple fatalities and extensive damage in Tarsin village and its surrounding agricultural lands following intense rainfall on 1 September 2025. Although these events pose a recurring threat to vulnerable communities, no previous landslide susceptibility assessment has been conducted in this region. This is largely due to the prolonged armed conflict in Darfur, which has restricted field-based investigations since 2003. In response to these challenges, the current study presents the first regional-scale rainfall-conditioned landslide susceptibility assessment for the Jebel Marra volcanic massif. The analysis integrates multi-source geospatial, geological, and geophysical datasets with a deep Convolutional Neural Network (CNN) model. A landslide inventory comprising 350 mapped events was developed using multi-temporal satellite imagery and visual interpretation. Key conditioning factors, including topographic parameters, hydrological characteristics, structural lineament density, vegetation cover derived from the Normalized Difference Vegetation Index (NDVI), and selected anthropogenic indicators, were incorporated into the model. The CNN model, trained and validated using stratified k-fold cross-validation, demonstrated high predictive capability (precision: 0.975, recall: 0.992, area under the curve (AUC): 1) and outperformed a benchmark Random Forest model. Feature importance analysis further indicates that elevation, curvature, and lineament density are the most influential conditioning factors controlling landslide occurrence in the study area. The resulting hazard map delineates high and very high hazard zones primarily concentrated in the central volcanic highlands and along major drainage corridors, representing approximately 15-20% of the study area. These findings provide a critical scientific basis for disaster risk reduction, humanitarian planning, and land-use management in this conflict-affected region.