Transformative Approaches in Landslide Detection-Synergizing Cutting-Edge Deep Learning Models with Optimal Fusion Strategies

Landslides are catastrophic natural disasters causing significant loss of life, displacement, and economic damage globally. To enhance landslide detection accuracy, an optimal methodology using advanced deep learning models has been developed. Five state-of-the-art models- VGG16, DenseNet201, ResNet50, InceptionV3, and a Generative Adversarial Network (GAN), were trained and evaluated. VGG16 and DenseNet201, the top-performing models, were integrated through feature fusion and global average pooling to combine their strengths. Data augmentation techniques were used to improve model robustness, and hyper-parameters were fine-tuned with a reduced learning rate for optimal performance. The combined model achieved 97.22% accuracy, 97.78% precision, 96.70% recall, and a 97.24% F1-score. Performance was demonstrated with detailed training and validation accuracy and loss curves, a confusion matrix, and an ROC curve with an AUC of 0.96. This methodology significantly enhances the reliability of emergency response efforts, particularly in disrupted communication networks, enabling timely interventions. A comparative analysis highlighted the superiority of this integrated approach over individual models. This study presents a novel pipeline for landslide detection, achieving advanced recognition through innovative model integration and optimization, contributing to effective disaster mitigation and long-term risk management.