Enhance the Accuracy of Landslide Detection in UAV Images Using an Improved Mask R-CNN Model: A Case Study of Sanming, China

Lu Yun, Xinxin Zhang, Yuchao Zheng, Dahan Wang, Lizhong HuaSensors 2023 · DOI

TL;DR

This paper introduces an enhanced Mask R-CNN model for more accurate landslide detection in high-resolution UAV imagery. By incorporating attention mechanisms, an improved feature pyramid network, and a guided anchor generation strategy, the model significantly boosts detection accuracy, addressing limitations of existing deep learning approaches in handling complex landslide features and background noise.

What Problem Does This Paper Address?

Landslides pose a significant threat globally, causing substantial human casualties and economic losses. Accurate and timely identification of landslide areas is crucial for disaster mitigation and risk assessment. Traditional methods, like ground surveys, are time-consuming and often incomplete. While remote sensing and machine learning have advanced landslide detection, existing deep learning models, particularly Mask R-CNN, struggle with challenges like background noise, varying landslide scales and shapes, and the loss of fine-grained feature information during processing. This can lead to missed detections or false positives, especially in complex terrains captured by high-resolution UAV imagery.

Methods

The authors propose an improved Mask R-CNN model by integrating three key enhancements:

  • CBAM Attention Mechanism: A Convolutional Block Attention Module (CBAM) is added to the ResNet backbone. CBAM combines channel and spatial attention to help the network focus on more relevant feature channels and spatial locations, suppressing irrelevant background information and better capturing landslide features.
  • Improved Feature Pyramid Network (FPN): A bottom-up channel is added to the FPN. This new path merges low-level (detailed) and high-level (semantic) feature maps, allowing the network to better integrate information across different scales and improve the handling of targets of varying sizes.
  • Guided Anchoring Region Proposal Network (GA-RPN): The traditional Region Proposal Network (RPN) is replaced with GA-RPN. This module uses key points to predict anchor positions and shapes, guiding anchor generation to be more adaptive to landslide variability and reducing redundant anchors, thereby improving detection accuracy.

These three modules are fused into a single improved Mask R-CNN model.

Datasets

The study utilizes high-resolution UAV images from Sanming City, Fujian Province, China, a region prone to landslides due to its mountainous terrain. The dataset consists of images acquired on June 12, 2019, following a heavy rainfall event.

  • Data Preprocessing: Images were cropped to 256x256 pixels.
  • Landslide Classification: Landslides were classified into categories based on area: Very small (<200 m²), Small (200-2000 m²), Medium (2000-20,000 m²), Large (20,000-200,000 m²), Very large (200,000-2,000,000 m²), and Huge (>2,000,000 m²).
  • Dataset Split: The dataset was split into training (60%), validation (20%), and testing (20%) sets. Data augmentation techniques (flipping) were applied to expand the training set, resulting in a total of 4050 landslide samples for model training.
  • Ground Truth: Landslides were visually interpreted by experts and field-verified, achieving a 97% final interpretation accuracy, with 241 landslides mapped.

Key Insights and Ideas

  • Attention is Key for Complex Features: The CBAM module's ability to focus on specific spatial and channel information proved crucial in distinguishing landslide features from similar-looking background elements like vegetation, reducing false detections.
  • Multi-Scale Feature Fusion is Essential: The improved FPN, by incorporating bottom-up channels, effectively combined low-level spatial details with high-level semantic context. This is vital for accurately detecting landslides of diverse sizes and shapes.
  • Adaptive Anchors Improve Robustness: Replacing the fixed anchors of the standard RPN with GA-RPN's guided, adaptive approach significantly improved the model's ability to handle the inherent variability in landslide appearance and scale, a common challenge in real-world imagery.
  • Fusion Outperforms Individual Modules: While each individual improvement (CBAM, improved FPN, GA-RPN) offered benefits, their synergistic fusion into a single model yielded the most substantial gains in landslide detection accuracy.

Key Results

The proposed improved Mask R-CNN model demonstrated significant performance improvements over the original Mask R-CNN:

  • Overall Accuracy: The improved model achieved a Precision of 93.9%, Recall of 91.4%, Accuracy of 92.6%, F1 score of 92.6%, and MIoU of 86.4%.
  • Improvements: These metrics represent an increase of 7% (Precision), 12.9% (Recall), 10.9% (Accuracy), 10.1% (F1), and 16.2% (MIoU) compared to the original Mask R-CNN.
  • Small Landslide Detection: The model showed particularly high accuracy (96.29%) in detecting small landslides, indicating its sensitivity to smaller-scale events.
  • Model Speed: The improved model achieved high accuracy with a minimal increase in detection speed (0.2493 s/iteration vs. 0.1660 s/iteration for the original Mask R-CNN), suggesting a good trade-off between accuracy and efficiency.
  • Comparison with Other Models: The improved model outperformed other common deep learning models like Faster R-CNN and SSD in terms of accuracy.

Limitations and Future Work

  • Background Noise: Despite improvements, the paper acknowledges that interference from background noise in surrounding features cannot be completely eliminated, especially when using only visible band imagery.
  • Generalizability: The study focused on a specific region (Sanming City) and rainfall-induced landslides. The authors suggest exploring the model's generality and portability by expanding the dataset with different data sources and geographical areas.
  • Data Fusion: Incorporating additional data sources, such as Digital Elevation Models (DEM), could further enhance landslide detection accuracy.
  • Parameter Optimization: The paper mentions that the depth and parameters of deep learning networks significantly influence accuracy. Future work could involve exploring optimal parameter combinations and data enhancement techniques.
  • Small Landslide Misses: While good at detecting small landslides, some instances were missed, potentially due to their extremely small size or obscuration by shadows or surrounding structures. The authors note that the relatively small number of landslide samples around buildings might have hindered learning.

Summary updated

BibTeX
@article{yun2023enhance,
  author = {Lu Yun and Xinxin Zhang and Yuchao Zheng and Dahan Wang and Lizhong Hua},
  title = {{Enhance the Accuracy of Landslide Detection in UAV Images Using an Improved Mask R-CNN Model: A Case Study of Sanming, China}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23094287},
}
Figure