Applying deep learning model to aerial image for landslide anomaly detection through optimizing process
Taiwan’s mountainous terrain is highly susceptible to landslides due to extreme weather events and anthropogenic activities. This study proposed a process offering an efficient reliable approach for rapid post-hazard landslide anomaly detection. The process employing the GANomaly deep learning model...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-12-01
|
| Series: | Geomatics, Natural Hazards & Risk |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/19475705.2025.2453072 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1832592109485424640 |
|---|---|
| author | Chwen-Huan Wang Li Fang Chiung-Yun Hu |
| author_facet | Chwen-Huan Wang Li Fang Chiung-Yun Hu |
| author_sort | Chwen-Huan Wang |
| collection | DOAJ |
| description | Taiwan’s mountainous terrain is highly susceptible to landslides due to extreme weather events and anthropogenic activities. This study proposed a process offering an efficient reliable approach for rapid post-hazard landslide anomaly detection. The process employing the GANomaly deep learning model to enhance landslide anomaly detection using high-resolution (25 cm) aerial imagery. The methodology encompasses multiple stages: pre-processing with RGB and LAB color corrections to improve image quality, slicing images into 128 × 128-pixel tiles, and applying augmentation technique by rotating tiles. These steps resulted in a dataset comprising approximately 505,000 normal tiles and 17,000 abnormal tiles, categorized into features including trees, roads, buildings, rivers, riverbanks, agricultural land, and landslide anomalies. Three GANomaly models were trained and tested using varying classification ratios, with datasets partitioned into training sets (normal images) and testing sets (normal and abnormal images). Model evaluation was conducted using confusion matrix parameters, with thresholds optimized through a weighted approach combining Youden’s index and the Closest method. Among the models, Train 2, which incorporated a 50% tree ratio and an average optimized threshold of 0.0124 (Closest method), achieved the highest AUC-ROC (∼0.98). Validation using pre- and post-Typhoon Morakot imagery demonstrated Train 2’s superior performance in accurately capturing landslide regions. |
| format | Article |
| id | doaj-art-7ba07f4cfd13498d988258ff7213f3a5 |
| institution | Kabale University |
| issn | 1947-5705 1947-5713 |
| language | English |
| publishDate | 2025-12-01 |
| publisher | Taylor & Francis Group |
| record_format | Article |
| series | Geomatics, Natural Hazards & Risk |
| spelling | doaj-art-7ba07f4cfd13498d988258ff7213f3a52025-01-21T15:57:48ZengTaylor & Francis GroupGeomatics, Natural Hazards & Risk1947-57051947-57132025-12-0116110.1080/19475705.2025.2453072Applying deep learning model to aerial image for landslide anomaly detection through optimizing processChwen-Huan Wang0Li Fang1Chiung-Yun Hu2Deparmentt of Civil Engineering, Chung Yuan Christian University, Taoyuan City, Taiwan (R.O.C.)School of Civil Engineering, FuJian University of Technology, Fuzhou, People’s Republic of ChinaDeparmentt of Civil Engineering, Chung Yuan Christian University, Taoyuan City, Taiwan (R.O.C.)Taiwan’s mountainous terrain is highly susceptible to landslides due to extreme weather events and anthropogenic activities. This study proposed a process offering an efficient reliable approach for rapid post-hazard landslide anomaly detection. The process employing the GANomaly deep learning model to enhance landslide anomaly detection using high-resolution (25 cm) aerial imagery. The methodology encompasses multiple stages: pre-processing with RGB and LAB color corrections to improve image quality, slicing images into 128 × 128-pixel tiles, and applying augmentation technique by rotating tiles. These steps resulted in a dataset comprising approximately 505,000 normal tiles and 17,000 abnormal tiles, categorized into features including trees, roads, buildings, rivers, riverbanks, agricultural land, and landslide anomalies. Three GANomaly models were trained and tested using varying classification ratios, with datasets partitioned into training sets (normal images) and testing sets (normal and abnormal images). Model evaluation was conducted using confusion matrix parameters, with thresholds optimized through a weighted approach combining Youden’s index and the Closest method. Among the models, Train 2, which incorporated a 50% tree ratio and an average optimized threshold of 0.0124 (Closest method), achieved the highest AUC-ROC (∼0.98). Validation using pre- and post-Typhoon Morakot imagery demonstrated Train 2’s superior performance in accurately capturing landslide regions.https://www.tandfonline.com/doi/10.1080/19475705.2025.2453072Landslide anomaly detectiondeep learningaerial imageimage pre-processingthreshold optimization |
| spellingShingle | Chwen-Huan Wang Li Fang Chiung-Yun Hu Applying deep learning model to aerial image for landslide anomaly detection through optimizing process Geomatics, Natural Hazards & Risk Landslide anomaly detection deep learning aerial image image pre-processing threshold optimization |
| title | Applying deep learning model to aerial image for landslide anomaly detection through optimizing process |
| title_full | Applying deep learning model to aerial image for landslide anomaly detection through optimizing process |
| title_fullStr | Applying deep learning model to aerial image for landslide anomaly detection through optimizing process |
| title_full_unstemmed | Applying deep learning model to aerial image for landslide anomaly detection through optimizing process |
| title_short | Applying deep learning model to aerial image for landslide anomaly detection through optimizing process |
| title_sort | applying deep learning model to aerial image for landslide anomaly detection through optimizing process |
| topic | Landslide anomaly detection deep learning aerial image image pre-processing threshold optimization |
| url | https://www.tandfonline.com/doi/10.1080/19475705.2025.2453072 |
| work_keys_str_mv | AT chwenhuanwang applyingdeeplearningmodeltoaerialimageforlandslideanomalydetectionthroughoptimizingprocess AT lifang applyingdeeplearningmodeltoaerialimageforlandslideanomalydetectionthroughoptimizingprocess AT chiungyunhu applyingdeeplearningmodeltoaerialimageforlandslideanomalydetectionthroughoptimizingprocess |