.png){kind=icon}
.png){kind=icon}
Return
An improved YOLO-based model for detecting small landslide deposits along mountainous rural roads via UAV imagery
.png){kind=icon}
OA
AI
J
S
Y
Q
Z
Y
T
X
DOI：10.1016/j.srs.2026.100444.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Landslide deposits distributed along mountainous roads are typically small in size, irregular in shape, and embedded within visually complex surroundings. These characteristics significantly hinder consistent and reliable detection, even when using high-resolution Unmanned Aerial Vehicle (UAV) imagery. To address these challenges, this study presents a novel recognition approach specifically tailored for landslide deposits in mountainous road environments. The method builds upon an enhanced and lightweight YOLOv8 framework, in which two key modules are integrated: the Context Aggregate (CA) module for capturing broad semantic contexts, and the Cross-Stage Partial Fusion-SwiftFormer Attention (C2F-SA) module, which incorporates an Inverted Residual-SwiftFormer (IR-SF) architecture combined with an Efficient Additive Attention (EAA) mechanism. Experiments were conducted using a dataset constructed from multi-temporal UAV imagery over mountainous road regions. The proposed YOLOv8-CAT-SF model achieved a mean Average Precision at an Intersection over Union threshold of 0.5 (mAP@0.5 ) of 0.7991 on the single test set, outperforming the original YOLOv8 model, which achieved a mAP@0.5 of 0.7655. In plain terms, mAP@0.5 evaluates whether landslide deposits are correctly identified and whether the predicted detection boundaries spatially overlap sufficiently with the ground-truth annotations. Five-fold cross-validation further yielded a mean mAP@0.5 of 0.7746 ± 0.0324 and a mAP@0.5– 0.95 of 0.3470 ± 0.0194, demonstrating stable generalization performance under different data partitions. The model also achieved a Precision of 0.7871, Recall of 0.7134, and pixel-level IoU of 0.6877. In terms of deployment efficiency, the proposed model contains 7.37 million parameters, requires 12.14 GFLOPs, has a model size of 14.43 MB, and reaches an average inference speed of 1.87 FPS on an Orange Pi 5 Pro edge device. These results indicate that the proposed framework improves the detection of small and fragmented landslide deposits while maintaining a lightweight structure suitable for UAV-based rapid inspection and potential edge deployment. Overall, the proposed method provides a practical and scalable solution for intelligent monitoring and risk management of landslide hazards along mountainous roads.
Keywords:
Landslide deposits
UAV imagery
Mountainous roads
Lightweight deep learning
YOLOv8
Remote sensing
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
5.2
Papers:
457
Citations:
980
Organization
Cited Papers