<p>Landslide detection is one of the key challenging tasks for disaster risk reduction and developing early warning systems. The conventional deep learning models such as U-Net and YOLO often fail to capture fine-scale features in heterogeneous satellite imagery. To overcome these limitations, we propose a Fusion-Aware Unified Framework that integrates deep fusion techniques such as Early Fusion and Late Fusion with advanced deep learning models such as attention mechanisms. Two datasets were used throughout this study, namely Landslide4Sense and HR-GLDD. We began by implementing basic U-Net and YOLO integrated with deep fusion techniques independently to identify landslides detection limitations. This approach established a baseline performance for comparison with more advanced models. Based on these results, two attention-driven fusion models were developed namely Attention-Driven Early Fusion model and Attention-Driven Late Fusion model. Following the evaluation of these models, we further extended the work by incorporating a self-attention mechanism with fusion techniques. We proposed two self-attention-based architectures named as SAMEL (Self-Attention Mechanism with Early Fusion), and SAMSNet (Self-Attention Mechanism with Stacking Network – A Late Fusion Technique) respectively. These models uses a self-attention module to effectively capture long-range spatial dependencies for landslide delineation. An ablation study was performed on different loss functions to evaluate their impact on models’ performance. Experimental results demonstrates that self-attention-guided models are effective in capturing accurate landslides. SAMEL achieved an accuracy of 98.47%, outperforming Attention-Driven Early fusion model (96.70%). On the other hand, SAMSNet achieved 98.5% accuracy, outperforming the Attention-driven Late Fusion model (97.48%). These results demonstrates the effectiveness of the proposed framework, offers a robust solution in timely disaster response and risk management.</p>