Deep Learning-Based Spatial Prediction of Landslide Risk in Coastal Areas Using GIS and Multicriteria Decision Making: A DeepLabV3+ Approach

Sustainable land-use planning and catastrophe risk reduction depend critically on landslide susceptibility mapping. The complex, nonlinear interconnections of environmental and human elements cause terrain instability and challenge conventional prediction methods. In this work, we offer a DeepLabV3+-based deep learning framework coupled with Geographic Information Systems (GIS) and Multi-Criteria Decision Making (MCDM) methods for spatial prediction of landslide risk, over the Dubai coastal and urban region (covering approximately 4,000 km²). The approach uses an annotated dataset for semantic segmentation and high-resolution satellite images from the Mohammed Bin Rashid Space Center (MBRSC). On Google Colab with GPU acceleration, the model is trained and verified and then further improved for computational efficiency on a Mac M1 machine. Our results show an Overall Accuracy (OA) of 91.3%, a Mean Intersection over Union (IoU) of 82.5%, and an F1-score of 88.4%, demonstrating strong classification performance throughout a range of land cover types. The Confusion Matrix analysis highlights strong segmentation accuracy for water bodies (94.2%) and structures (92.4%). Considerable misclassification between roadways and unpaved terrain results from spectral similarities. Furthermore, the perclass Dice Coefficient analysis confirms that the model can efficiently discriminate intricate topographical patterns. Especially in fast-expanding areas like Dubai, UAE, it provides a scalable solution for landslide susceptibility mapping, catastrophe risk management, and sustainable urban design. Future work will explore multi-sensor data fusion, real-time inference, and applying explainable artificial intelligence (XAI) techniques to enhance model interpretability in dynamic terrain settings.