A Hybrid Deep Learning and Object Based Image Analysis Framework for Enhanced Landslide Detection

Landslide detection is crucial for managing human settlements and infrastructure-related issues. This study aims to enhance landslide detection accuracy by integrating a deep learning (DL) model with object-based image analysis (OBIA) in Bijie. The ResU-net model is developed for semantic segmentation, trained, and tested on TripleSat imagery, enabling the identification of landslide-related patterns and features. Additionally, a rule-based OBIA approach is implemented to capture distinctive landslide characteristics. The OBIA analysis is applied to the original image dataset and combined with the ResU-net heatmap to assess accuracy. By leveraging the strengths of DL and OBIA, the ResU-net-OBIA approach demonstrates improved landslide mapping performance. The evaluation involves a manually digitised image inventory comprising 770 landslide and 2003 non-landslide images, covering an area of 26,322 square kilometres from May to August 2018. The study underscores the significance of incorporating distinct features and utilising both DL and OBIA techniques. The proposed framework offers a positive direction for future research, enhancing landslide monitoring, mitigation efforts, and disaster management strategies. The integration of DL and OBIA bridges the gap between pixel-based and object-based analysis, facilitating a comprehensive understanding of landslide patterns and features. The study showcases improved accuracy in landslide identification and differentiation from non-landslide areas. The ResU-net-OBIA framework holds promise for effective landslide detection and management in Bijie, and other regions grappling with landslide-related challenges. In conclusion, this study highlights the potential of integrating deep learning and object-based image analysis to enhance landslide detection accuracy, contributing to improved landslide mapping and informed decision-making in disaster management.