Enhancing landslide anomaly detection in aerial imagery through pre-processing and GANomaly deployment
Enhancing landslide anomaly detection in aerial imagery through pre-processing and GANomaly deployment
The application of aerial imagery for landslide anomaly detection encapsulates a promising avenue for early geohazard identification. This study utilizes GANomaly, a model built upon Generative Adversarial Networks (GAN), to identify anomalies by learning to reconstruct normal data and measuring reconstruction errors to discern anomalies. Initially, a pre-processing process was developed to enhance the clarity and quality of images, thereby facilitating more effective training and testing by GANomaly. This pre-processing process encompasses several steps: image RGB channels adjustments, image slicing into tiles, tile classifications, and tile rotations. The GANomaly model is then employed, trained on a substantial dataset derived from the pre-processing stage, and tested to evaluate its efficacy in detecting landslide anomalies within the aerial imagery. The entire process involved the selection and analysis of over 120,000 tiles, each acting as a data point to feed the robust anomaly detection framework. The results highlight the potential of integrating advanced pre-processing techniques with GAN-based anomaly detection models, forging a framework for real-time geohazard monitoring and early-warning systems. The results were verified from the confusion matrix parameters to ensure the potential of integrating advanced pre-processing techniques with GAN-based anomaly detection models.
C. H. Wang; K. C. Lo; C. Y. Hu; P. H. Wang
18th European Conference on Soil Mechanics and Geotechnical Engineering (ECSMGE2024)
B - Geohazards
Contact Information
- Secretary General - Dr A M McNamara
- Email: [email protected]
- Phone: +44 20 7040 8154
- City, University of London
- Northampton Square
- London EC1V 0HB. United Kingdom