Automated Landslides Detection for Mountain Cities Using Multi-Temporal Remote Sensing Imagery

Zhong Chen, Yifei Zhang, Chao Ouyang, Feng Zhang, Jie Ma2018 · DOI

TL;DR

This paper introduces an automated method for detecting landslides in mountain cities using multi-temporal remote sensing imagery. It leverages a Deep Convolutional Neural Network (DCNN) for change detection and a Spatial Temporal Context Learning (STCL) approach to accurately identify landslide areas, outperforming previous methods by reducing commission errors and improving detection rates.

What Problem Does This Paper Address?

Landslides in mountainous urban areas pose a significant threat, causing substantial casualties and economic losses. Accurately and quickly identifying these landslide-prone areas is crucial for disaster response and risk assessment. Traditional methods often struggle due to the complex appearance of landslides, making it difficult to find consistent spatial regularities. Relying solely on spatial information or artificial features leads to poor performance, while existing multi-temporal approaches can be prone to omission and commission errors, especially when dealing with diverse land cover changes. Furthermore, many methods require manual intervention or pre-existing landslide inventories, which are not always available. This paper aims to address these limitations by proposing an automated, knowledge-based approach that utilizes multi-temporal remote sensing data to detect landslides in mountain cities.

Methods

The proposed method is a multi-stage process designed for automated landslide detection:

  1. Irrelevant Area Removal: Initially, areas that are unlikely to be landslides are removed. This includes cloud-covered regions, as well as areas with vegetation, water, and built-up land in the post-event imagery. This step is crucial for reducing computational load and focusing the analysis on potentially relevant areas. Vegetation and water are identified using Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI), respectively, while built-up areas are detected using the PANTEX index.
  2. Change Detection using DCNN: A Deep Convolutional Neural Network (DCNN) is employed to identify areas where significant changes have occurred between pre- and post-event images. The DCNN extracts deep features from image patches, and the Manhattan distance between these features from the two time points is used to quantify the change. A threshold is then applied to distinguish changed areas from unchanged ones. This approach leverages the DCNN's ability to learn robust representations of image content.
  3. Spatial Temporal Context Analysis (STCL): For the identified changed areas, a Spatial Temporal Context Learning (STCL) approach is used to pinpoint actual landslide areas. STCL utilizes both spatial and temporal information, focusing on changes in NDVI and PANTEX. It models the spatial context of changed pixels and uses this information to determine if the observed changes are consistent with landslide phenomena (e.g., disappearance of vegetation and buildings). This step helps to filter out changes caused by non-landslide events.
  4. Post-Processing: Finally, a post-processing step refines the detected landslide areas. This involves using slope information derived from a Digital Elevation Model (DEM) to confirm plausibility (average slope > 7°). Small, scattered detected regions are filtered out to reduce false alarms. Additionally, changes in DEM are considered to include landslide areas that might not exhibit significant cover type changes but have undergone topographic shifts.

Datasets

The proposed method was validated using multi-temporal remote sensing imagery from three Chinese cities:

  • Shenzhen: GF-1 data (24 December 2013 and 29 December 2015)
  • Zhouqu County: Landsat-8 data (13 July 2009 and 10 September 2010)
  • Beichuan County: Formosat-2 data (5 April 2006 and 16 May 2008)

All images used were 4-band (blue, green, red, and near-infrared) with a spatial resolution of 8m for GF-1 and 30m for Landsat-8. A DEM with 5m spatial resolution was also used. The DCNN model was trained on the SAT-4 dataset.

Key Insights and Ideas

  • Leveraging Land Cover Change Knowledge: The core idea is that landslides in urban mountain areas cause specific, detectable changes in land cover (vegetation, buildings). By focusing on these changes and filtering out other types of land cover variations, the method can achieve higher accuracy.
  • DCNN for Robust Feature Extraction: Using DCNN for change detection is a significant advancement. It moves beyond hand-crafted features to learn complex, multi-level representations of image content, making it more effective at capturing subtle changes indicative of landslides.
  • STCL for Contextual Discrimination: The STCL approach is key to distinguishing landslide-induced changes from other types of surface changes. By considering the spatial and temporal context of changes, it can effectively rule out false positives that might arise from non-landslide events.
  • Multi-Stage Filtering for Accuracy: The sequential application of irrelevant area removal, DCNN-based change detection, STCL, and post-processing creates a robust filtering mechanism that progressively refines the detection results, leading to lower commission errors.
  • DEM for Completeness: Incorporating DEM data, both for slope analysis and detecting topographic changes, helps to ensure that the detection is more complete, accounting for landslides that might not have obvious surface cover changes.

Key Results

The proposed method demonstrated strong performance across the three validation sites:

  • Landslide Areal Extent:
    • Shenzhen: DP 79.9%, QP 73.4%, CE 9.9%
    • Zhouqu County: DP 70.3%, QP 61.1%, CE 17.6%
    • Beichuan County: DP 88.5%, QP 80.1%, CE 9.6%
    • Overall, the detection percentage (DP) for landslide areal extent was over 70.3%, with a commission error (CE) below 17.6%.
  • Landslide Areas:
    • Shenzhen: DP 100%, QP 71.4%, CE 28.6%
    • Zhouqu County: DP 100%, QP 66.7%, CE 33.3%
    • Beichuan County: DP 81%, QP 77.4%, CE 5.9%
    • The method achieved a 100% recall rate for landslide areas in Shenzhen and Zhouqu County.
  • Qualitative Results: The method successfully identified large landslides and reduced commission errors by filtering irrelevant areas. The post-processing step helped to refine the shape and outline of detected landslides.

Limitations and Future Work

  • Limitations:
    • The method might miss landslide areas that do not conform to the assumed land cover change knowledge, such as debris flow source areas where erosion is dominant or rivers buried by landslides.
  • The spatial resolution of the remote sensing images used (8m and 30m) might limit the detection of very small or subtle landslides.
  • Future Work:
    • Develop more comprehensive models to address landslide types that do not fit the current land cover change assumptions.
  • Investigate the use of higher spatial resolution remote sensing imagery to improve the detection of smaller landslides.
  • Explore extending the approach to cover larger geographical areas and more complex natural environments.

Summary updated

BibTeX
@misc{chen2018automated,
  author = {Zhong Chen and Yifei Zhang and Chao Ouyang and Feng Zhang and Jie Ma},
  title = {{Automated Landslides Detection for Mountain Cities Using Multi-Temporal Remote Sensing Imagery}},
  year = {2018},
  doi = {10.3390/s18030821},
}
Figure