A Research on Landslides Automatic Extraction Model Based on the Improved Mask R-CNN

Peng Liu, Yongming Wei, Qinjun Wang, Jingjing Xie, Yu Chen, Zhichao Li, Hongying ZhouISPRS Int. J. Geo Inf 2021 · DOI

TL;DR

This paper introduces an improved Mask R-CNN model for automatically extracting landslides from remote sensing imagery, crucial for disaster response. By enhancing the feature extraction, incorporating multi-scale information, and refining boundary detection, the proposed model significantly outperforms the standard Mask R-CNN in accuracy and efficiency.

What Problem Does This Paper Address?

Landslides are destructive secondary geological hazards, especially after earthquakes. Rapidly identifying their location and extent is vital for effective disaster relief and rescue operations. Traditional methods like visual interpretation are time-consuming and subjective, while automated methods often struggle with the diverse characteristics of landslides (shape, size, texture) and the complexity of remote sensing data. Existing deep learning approaches, while promising, still face challenges in accurately delineating landslide boundaries and handling small-sized landslides, leading to potential delays in providing critical information for emergency response.

Methods

The core of this research is an enhanced Mask R-CNN architecture designed specifically for landslide extraction. The key improvements include:

  • ResNeXt Feature Extractor: Replacing the standard ResNet backbone with ResNeXt (specifically ResNeXt50) for more efficient and effective extraction of landslide features. ResNeXt's "cardinality" approach allows for better feature representation without a significant increase in parameters.
  • Improved Feature Pyramid Network (FPN): Introducing bottom-up channels in the FPN. This allows for better integration of low-level positional information with high-level semantic information, crucial for detecting smaller landslides and improving overall feature representation across scales.
  • Edge Loss Function: Adding an "edge loss" to the overall loss function. This auxiliary loss specifically penalizes discrepancies between predicted and actual landslide edges, leading to more accurate boundary detection and a more precise segmentation mask.

Datasets

The study utilizes Unmanned Aerial Vehicle (UAV) remote sensing data from the Jiuzhaigou area in Sichuan Province, China.

  • Training Set: 366 landslides covering an area of 34.6 km².
  • Test Set: 233 landslides covering an area of 12.6 km².

The landslide interpretation map used for ground truth was field-verified to achieve 98% accuracy. Data augmentation techniques (rotations and flips) were applied to increase the training set size to 9762 samples.

Key Insights and Ideas

  • ResNeXt's Efficiency: ResNeXt-50 offers comparable accuracy to ResNet-101 but with half the computational cost, making it a more efficient choice for feature extraction.
  • FPN's Multi-Scale Power: The improved FPN, with its bottom-up path, effectively bridges the gap between low-level spatial details and high-level semantic understanding, which is critical for detecting landslides of varying sizes.
  • Edge Loss for Precision: Explicitly incorporating edge information into the loss function directly addresses the challenge of accurately delineating landslide boundaries, a common weakness in segmentation tasks. This not only improves accuracy but also accelerates convergence during training.
  • End-to-End Solution: The proposed model offers an end-to-end solution that can directly process post-earthquake imagery, avoiding the need for pre-earthquake data which may not always be available.

Key Results

The improved model, named ResNeXt50+Improved FPN+Ledge, demonstrated significant performance gains on the Jiuzhaigou earthquake landslide dataset:

  • Precision: 95.8%
  • Recall: 93.1%
  • Overall Accuracy (OA): 94.7%
  • Mean Intersection over Union (mIoU): 89.6%
  • F1 Score: 94.5%

Compared to the original Mask R-CNN (represented by ResNeXt50 in their experiments), these metrics show substantial improvements:

  • Precision improved by 13.9%.
  • Recall improved by 13.4%.
  • OA improved by 9.9%.

The paper also highlights that the improved model significantly reduced missed extractions and improved edge detection accuracy compared to baseline models.

Limitations and Future Work

  • Limited Training Data: The authors acknowledge that the training set of seismic landslides is still limited, leading to some residual errors in extraction.
  • Generalization: To make the model more practical, future work should focus on extending the training set to include remote sensing images of different types and resolutions, as well as landslides from various geographical regions and types. This will improve the model's generalization capabilities.

Summary updated

BibTeX
@article{liu2021research,
  author = {Peng Liu and Yongming Wei and Qinjun Wang and Jingjing Xie and Yu Chen and Zhichao Li and Hongying Zhou},
  title = {{A Research on Landslides Automatic Extraction Model Based on the Improved Mask R-CNN}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2021},
  doi = {10.3390/ijgi10030168},
}
Figure