<p>Landslides are destructive geological phenomena that occur when soil, rock, and debris slide down slopes due to factors such as heavy rainfall, seismic activity, or human intervention. They cause severe damage to infrastructure, ecosystems, and human life, particularly in Karnataka’s Western Ghats, which are highly vulnerable to steep terrains and intense monsoonal rainfall. This study presents GeoLSTM-FuzzNet (Geographical Long Short-Term Memory–Fuzzy Neural Network), a deep learning framework designed to enhance landslide susceptibility prediction by effectively integrating geographical and temporal information. The model consists of three core components: the Landslide Index Fuzzy Inference System (LD-FIS), which applies fuzzified geographic features such as elevation, slope, and Normalized Difference Vegetation Index (NDVI) to handle spatial uncertainty; the Spatio-Temporal Prediction Network (ST-DPN), which utilizes Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) structures to analyze temporal rainfall sequences; and the Adaptive Feature Fusion Unit (AFFU), which employs learnable fusion weights to dynamically balance spatial and temporal contributions. The framework was trained on a comprehensive multi-source dataset from Karnataka comprising a Digital Elevation Model (DEM), Normalized Difference Vegetation Index (NDVI), rainfall data, and verified landslide inventories. The experimental results show that GeoLSTM-FuzzNet significantly outperforms the baseline CNN model, achieving 92% accuracy, 91% precision, 89% recall, an F1-score of 0.90, and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.94, compared with 81% accuracy and 0.80 AUC-ROC for the CNN model. These findings demonstrate an 11% improvement in accuracy and a 14% increase in AUC-ROC, confirming the robustness and reliability of GeoLSTM-FuzzNet for real-time early warning systems and disaster risk management in landslide-prone regions.</p>