Landslide Detection and Mapping Using Deep Learning Across Multi-Source Satellite Data and Geographic Regions

The selected study areas represent diverse geographic and climatic conditions. Their locations are depicted on a global landslide susceptibility map generated using multiple explanatory variables such as slope degree, forest loss, geology, road networks, and fault lines.

1. Iburi-Tobu Area of Hokkaido, Japan:

• Hit by a magnitude 6.6 earthquake on September 6, 2018, triggering over 5600 landslides.
• Landslides were exacerbated by preceding heavy rainfall from Typhoon Jebi.
• Landslide inventories were created using very high-resolution aerial images.

2. Kodagu District of Karnataka, India:

• Experienced extreme rainfall in August 2018, triggering severe landslides and flash floods.
• Landslides were linked to deforestation, unplanned urbanization, and mining activities.
• Previous studies have applied unsupervised learning techniques for landslide detection.

3. Rasuwa District of Bagmati, Nepal:

• One of the most landslide-prone regions in the Himalayas.
• Major landslides occurred due to the 2015 Gorkha and Dolakha earthquakes.
• Landslide inventory compiled from GPS field surveys and visual interpretation of high-resolution images.

4. Western Taitung County, Taiwan:

• Landslides frequently triggered by typhoons and earthquakes.
• Typhoon Morakot (2009) caused extensive landslides, destroying villages and infrastructure.
• Landslide inventory derived from previous studies and Google Earth images.

1. Sentinel-2:

• Provides multi-spectral imagery with 13 bands at spatial resolutions of 10, 20, and 60 meters.
• High revisit frequency (2–3 days at mid-latitudes) enables continuous monitoring.
• Sentinel-2 data were obtained from Google Earth Engine (GEE), ensuring cloud-free imagery for analysis.

2. ALOS PALSAR:

• Provides synthetic aperture radar (SAR) data with a 12.5m spatial resolution.
• The DEM and slope layers derived from ALOS PALSAR were used to supplement optical imagery.

To ensure high-quality landslide annotations, we employed a two-step workflow:

1. Object-Based Image Analysis (OBIA):

• Computed image difference indices from pre- and post-landslide images.
• Performed multi-resolution segmentation and rule-based classification.

2. Manual Verification:

• Used high-resolution Google Earth imagery and existing landslide inventories.
• Ensured accurate delineation of landslide boundaries.

The Landslide4Sense benchmark dataset includes 128×128 window-size patches, each containing 14 distinct data layers. The first 12 bands consist of multi-spectral data from Sentinel-2, while bands 13 and 14 represent the Digital Elevation Model (DEM) and slope data derived from ALOS PALSAR. Each patch is accurately labeled, with ground truth polygons outlined in red to indicate landslide areas.

Visualizing each unique layer inside the generated landslide dataset's 128x128 window-size patches. The first 12 bands feature multi-spectral data from Sentinel-2, while bands 13 and 14 contain DEM data and slope from ALOS PALSAR. The patches in the last column are accurately labeled, and they are complimented by red polygons signifying the landslide category. These patches in last column refers to the ground truth Polygons.

The 14 layers in the Landslide4Sense dataset:

• Sentinel-2 band 1: Blue spectral band data
• Sentinel-2 band 2: Green spectral band data
• Sentinel-2 band 3: Red spectral band data
• Sentinel-2 band 4: Near Infrared (NIR) spectral band data
• Sentinel-2 band 5-7: Shortwave Infrared (SWIR) spectral band data
• Sentinel-2 band 8: NIR spectral band data
• Sentinel-2 band 9: Water Vapour (WV) spectral band data
• Sentinel-2 band 10: Cirrus (CI) spectral band data
• Sentinel-2 band 11-12: SWIR spectral band data
• Digital Elevation Model (DEM): Elevation information data
• Slope: Slope information data

Dataset Statistics:

• 3799 annotated image patches (128 × 128 pixels each, with a resolution of 10 meters per pixel)
• 14 bands per image patch
• Dataset split: 959 patches for training, 2840 patches for testing

Each patch contains pixel-wise labels indicating landslide and non-landslide areas. The dataset exhibits significant variability in landslide shape, size, distribution, and frequency across study areas. Sentinel-2 bands 4 and 5 show the highest spectral differences between landslide and non-landslide areas.

Our results demonstrate that ResNet34, VGG-16, and EfficientNet-B0 achieved the highest F1 Scores, indicating superior performance in distinguishing landslide-prone areas from non-landslide regions. The ResNet34-based U-Net model attained the best balance between precision and recall, achieving an F1 Score of 0.7470, making it the most reliable among the tested architectures. Notably, VGG-16 and EfficientNet-B0 also performed well, with F1 Scores of 0.7357 and 0.7341, respectively.

The classic U-Net architecture, while still effective, demonstrated a lower F1 Score of 0.7012, highlighting the advantage of deeper and more advanced feature extraction architectures like ResNet and EfficientNet-B0. SeResNet-50 and SeResNeXt50_32x4d also showcased competitive performance, emphasizing the benefit of integrating Squeeze-and-Excitation modules for better feature representation.

The results indicate that hybrid models leveraging deeper feature extraction mechanisms significantly enhance landslide detection performance compared to standard U-Net. The ability to capture both local and global contextual information plays a crucial role in improving segmentation quality. This study further reinforces the need for models that efficiently integrate multi-scale feature representations for landslide susceptibility mapping in complex terrains.

Table: Comparison of performance evaluation metrics of segmentation models tested on the Landslide4Sense dataset.

Performance Metrics for Landslide Prediction in U-Net

Performance Metrics for Landslide Prediction in ResNet34, VGG-16, EfficientNet-B0, ResNeXt50_32X4D, SeResNet-50, DenseNet121, SeResNeXt50_32x4D, InceptionV4, InceptionResNetV2, DeepLabV3+, MobileNetV2, MiT-b1_14C.

Comparative Model Performance:

The comparative evaluation underscores the strength of ResNet34 and VGG-16 as encoders in the U-Net framework. These models not only offer high predictive accuracy but also minimize false detections. The increase in recall values across these models suggests a significant improvement in detecting subtle landslide regions, which is critical for real-world applications where missing a landslide event could lead to disastrous consequences. By leveraging advanced deep learning techniques, our study demonstrates that multi-source satellite imagery, when processed with optimized architectures, can significantly enhance landslide detection accuracy. These findings contribute to the growing field of deep learning applications in geospatial analysis, paving the way for more reliable and scalable landslide prediction systems.

RESULT

We evaluated various deep learning models for landslide detection using the Landslide4Sense dataset with Sentinel-2 imagery and ALOS PALSAR elevation data. The ResNet34-based U-Net achieved the highest F1 Score of 0.7470 with balanced precision (0.7737) and recall (0.7267). VGG16 and EfficientNet-B0 also performed well with F1 Scores of 0.7357 and 0.7341 respectively. Advanced architectures significantly outperformed the classic U-Net (F1: 0.7012), demonstrating the importance of deeper feature extraction mechanisms for complex geospatial data.

CONCLUSION

This study demonstrates that hybrid deep learning models with advanced feature extraction significantly outperform traditional U-Net for landslide detection. ResNet34-based U-Net emerged as the most reliable architecture with an F1 Score of 0.7470. Multi-source data integration combining optical imagery with elevation information proves crucial for accurate landslide identification. These findings contribute to developing more reliable disaster risk management and early warning systems for landslide-prone regions.