Automated detection of landslide using synergizing dual Graph Convolutional Networks, googlenet, and machine learning techniques

Abstract

This study explores a synergistic approach to automated landslide detection in Centro Fluminense, leveraging advanced deep learning and machine learning frameworks. The proposed methodology integrates dual Graph Convolutional Networks (DGCN) with GoogLeNet to analyze topographic and 175 pre-historic landslide data for precise mapping. A curated dataset of landslide and 195 topographic images underscores the novelty and effectiveness of this approach. The framework employs dual Graph Convolutional Networks to capture spatial dependencies and GoogLeNet to extract deep spatial features effectively. A machine learning model complements these networks to refine predictions through iterative learning processes. The study evaluates network modelling through DGCN and GoogLeNet, focusing on training and validation accuracy. Training datasets demonstrated consistent improvement in classification accuracy, increasing from 66% to 93%, while validation datasets achieved high precision, with values rising from 78% to 99%. The results emphasize the model's robustness and scalability in addressing spatial heterogeneity and complex topographic conditions. Performance metrics were rigorously analyzed, indicating a significant alignment with ground-truth data, as evidenced by a coefficient of determination (R²) of 0.92 and a mean absolute error (MAE) of 4%. The integration of DGCN and GoogLeNet outperformed conventional methods by capturing intricate spatial and contextual features. This comprehensive framework ensures reliable and automated detection, crucial for disaster risk management in regions prone to landslides. In addition to predictive modelling, the study highlights the role of preprocessing techniques, including hillside and LULC analysis, in enhancing detection capabilities. A comparative analysis of models reveals the superiority of the dual network approach over single-framework architectures, particularly in terms of accuracy and adaptability to diverse datasets. This study provides a novel contribution to landslide mapping by combining topographical insights with advanced network architectures. The proposed framework demonstrates the potential for deployment in other regions with similar geological settings, paving the way for improved disaster preparedness and management strategies.