Deep Learning Techniques for Identifying Slope Anomalies in Multi-Resolution Aerial Images

This study developed a deep learning-based approach for detecting slope anomalies in mountainous regions using multi-resolution aerial imagery and the GANomaly model. High-resolution datasets, including 25 cm resolution aerial aircraft imagery and 70 cm resolution satellite imagery, were utilized. Key pre-processing steps included color space correction, resolution conversion, image slicing, and data augmentation through tile rotation. Resolution conversion played an important role in integrating diverse image sources into a unified 25 cm format, ensuring consistency during training. Five training combinations with varying ratios of ASRS ($40 \%-80 \%$) and Satellogic ($20 \%-60 \%$) imagery were developed to train models capable of handling multi-resolution data. The testing phase evaluated three distinct groups: Group 1 used 25 cm imagery, Group 2 incorporated 70 cm imagery converted to 25 cm alongside 25 cm images, and Group 3 utilized original 70 cm imagery with 25 cm images converted to 70 cm. Results showed that Group 1 exhibited consistent trends but lower performance values, while Group 2 demonstrated moderate improvements with performance values of $\mathbf{0. 8 - 0. 9}$. Group $\mathbf{3}$ achieved the highest performance values, establishing an optimal $50 \%: 50 \%$ ratio of 25 cm and 70 cm imagery for anomaly detection. This approach provides a robust foundation for future landslide monitoring and disaster management in diverse terrains.