17:15 〜 19:15
[HTT15-P02] Review of Remote Sensing and Deep Learning Integration for Landslide Mapping: Trends, Challenges, and Future Directions
キーワード：斜面崩壊、マッピング手法、深層学習、リモートセンシング
Landslides, triggered by various hazards and predisposing factors, pose severe risks to human lives and infrastructure. Traditional mapping methods, such as field surveys and manual interpretation of aerial imagery, provide high accuracy but are time-consuming, labor-intensive, and highly dependent on expert knowledge. The advancement of remote sensing and deep learning has significantly improved landslide detection by enabling large-scale automated mapping. This study reviewed existing literature integrating remote sensing with deep learning for landslide mapping. A literature database was compiled from the Web of Science using Boolean search operators and keywords related to landslides and deep learning. First, we evaluated model trends and performance using F1 scores across four deep learning model types: CNN-based, transformer-based, one-stage object detection, and two-stage object detection. We then analyzed the spatial distribution of landslide inventories developed for deep learning and the regions where these models were applied. Additionally, we assessed the remote sensing platforms used in these studies to evaluate their applicability in AI-based landslide detection.
The review covered 254 related papers, revealing that since 2019, research on the integration of deep learning and remote sensing for landslide detection has surged, showing an increasingly rapid growth trend and demonstrating strong performance in landslide mapping. Among the four deep learning model types, CNN-based models (72 studies) remained dominant, while transformer-based models exhibited a sharp increase from 2023 (three studies) to 2024 (25 studies). In terms of performance, one-stage object detection achieved the highest F1 scores (81.50%), followed by transformers (76.06%), CNNs (74.11%), and two-stage models (72.59%). Despite their high performance, AI-based landslide inventories and model applications showed significant regional biases, with most studies concentrated in East and South Asia. China ranked first with 115 studies, followed by Japan with 39, while Africa and South America remained underrepresented. Regarding remote sensing platforms, optical imagery was the most widely used (127 studies), followed by multi-source datasets (75) and landslide inventories (66), while SAR remained underexplored (24).
Future research should focus on publishing more geographically diverse datasets for AI-based landslide detection, expanding deep learning model applications, and optimizing multi-source data integration. These advancements will enhance the reliability and scalability of landslide detection frameworks, contributing to global geohazard resilience.
The review covered 254 related papers, revealing that since 2019, research on the integration of deep learning and remote sensing for landslide detection has surged, showing an increasingly rapid growth trend and demonstrating strong performance in landslide mapping. Among the four deep learning model types, CNN-based models (72 studies) remained dominant, while transformer-based models exhibited a sharp increase from 2023 (three studies) to 2024 (25 studies). In terms of performance, one-stage object detection achieved the highest F1 scores (81.50%), followed by transformers (76.06%), CNNs (74.11%), and two-stage models (72.59%). Despite their high performance, AI-based landslide inventories and model applications showed significant regional biases, with most studies concentrated in East and South Asia. China ranked first with 115 studies, followed by Japan with 39, while Africa and South America remained underrepresented. Regarding remote sensing platforms, optical imagery was the most widely used (127 studies), followed by multi-source datasets (75) and landslide inventories (66), while SAR remained underexplored (24).
Future research should focus on publishing more geographically diverse datasets for AI-based landslide detection, expanding deep learning model applications, and optimizing multi-source data integration. These advancements will enhance the reliability and scalability of landslide detection frameworks, contributing to global geohazard resilience.