TL;DR
This paper introduces a novel semi-automatic method (SCDT) for regional landslide identification by integrating landslide susceptibility analysis with change detection. This approach significantly improves the accuracy of landslide identification compared to traditional change detection methods by effectively filtering out false positives caused by non-landslide changes.
What Problem Does This Paper Address?
Identifying landslides accurately and efficiently is crucial for hazard prevention and mitigation. Traditional methods like visual interpretation and field surveys are labor-intensive, time-consuming, and can be difficult in remote or rugged terrain. While remote sensing and GIS have advanced landslide mapping, current automatic or semi-automatic methods still struggle with accuracy and efficiency, often producing false positives due to various land cover changes. This paper aims to bridge this gap by developing a more robust and accurate semi-automatic landslide identification method.
Methods
The core of this research is the Susceptibility-based Change Detection Thresholding (SCDT) method. It combines two main components:
- Change Detection (CD): This step uses multi-temporal Landsat 8 images to identify areas where significant changes have occurred. A difference image (DI) is generated using Change Vector Analysis (CVA), and a multi-threshold method (based on mean and standard deviation of pixel change values) is applied to classify pixels into "landslide candidate regions," "uncertain regions," and "non-landslide candidate regions." This is an improvement over simpler binary thresholding methods by offering more nuanced classification.
- Landslide Susceptibility Analysis (LSA): This step uses a Random Forest (RF) model to predict the likelihood of landslides occurring in the study area. It incorporates 12 parameters related to topography, geomorphology, geology, and land cover. Crucially, landslide data from the period of interest (2014-2016) is excluded from the training data to ensure the susceptibility map reflects the potential for future landslides, not just past occurrences.
The SCDT method then integrates the results of CD and LSA. Candidate regions identified by CD are further refined using the susceptibility map. Pixels are classified into five classes based on their susceptibility and change detection values. A region is classified as a landslide if a certain percentage of its pixels fall into specific classes, effectively using susceptibility to filter out non-landslide changes identified by CD.
Datasets
- Landsat 8 OLI Images: Used for change detection. Images from May 2014 and May 2016 were utilized as pre- and post-event images.
- Landslide Inventory: Provided by the Jilin Institute of Geological Environment Monitoring (JIGEM). This dataset documented landslide locations and boundaries between 2000 and 2016. For this study, landslides occurring between 2014 and 2016 were specifically used for validation and analysis.
- Topographic and Geomorphological Data: Derived from a 1:50,000 topographic map (2004) and other sources, including slope gradient, altitude, curvature, terrain roughness index, etc.
- Geological and Land-Use/Land-Cover (LULC) Data: Provided by JIGEM.
Key Insights and Ideas
- Synergy of CD and LSA: The paper demonstrates that combining change detection with landslide susceptibility analysis is more effective than either method alone. Susceptibility analysis acts as a powerful filter, reducing false positives generated by change detection due to non-landslide phenomena (e.g., vegetation changes, urban development).
- Strategic Exclusion of Recent Landslide Data in LSA: By excluding recent landslide data (2014-2016) from the LSA training set, the model is trained to predict the potential for landslides, rather than simply identifying areas that have already experienced them. This is crucial for a predictive identification method.
- Multi-Thresholding for CD: Employing a multi-threshold method for change detection provides a more nuanced classification of change areas compared to simpler binary methods, allowing for better initial discrimination of potential landslide areas.
- Pixel-Based Approach with LSA Filtering: While CD is pixel-based, the integration with LSA allows for a more robust identification of actual landslide areas by considering broader environmental factors.
Key Results
- Improved Correctness: The SCDT method showed a significant improvement in correctness across the entire study area and three sub-areas. For the whole study area, correctness increased by 29.60% compared to masked CDT. In sub-areas A, B, and C, improvements were 27.90%, 23.99%, and 4.57%, respectively.
- Identical Completeness: The completeness of SCDT was the same as CDT, indicating that both methods identified a similar proportion of actual landslide areas within the reference boundaries. This is attributed to the pixel-based nature of the underlying CD.
- Slightly Improved Quality: The quality index, which considers both correctness and completeness, showed a slight improvement with SCDT, ranging from 0.12% to 12.03% across different areas.
- Effective Elimination of False Positives: Qualitative evaluations in sub-areas A and B demonstrated that SCDT successfully eliminated many candidate regions caused by farmland changes, road construction, and vegetation cover variations, which were misidentified by CDT alone.
Limitations and Future Work
- Focus on Specific Landslide Types: The study primarily used rotational and planar shallow landslide data for susceptibility analysis. The method may not perform as well for other landslide types like debris flows, which were sometimes misidentified as actual landslides.
- LULC Data Accuracy: Potential errors in the LULC data could lead to misidentification.
- Reference Boundary Inclusions: The reference landslide boundaries sometimes included non-trigger areas, which could have reduced the accuracy of the LSA.
- Low Resolution Data: The use of 10m resolution Landsat data limits the representation of real topography and detailed landslide features.
- Future Work Suggestions:
- Combine SCDT with object-based image analysis for improved boundary detection.
- Utilize higher-resolution imagery (e.g., 10m pixel resolution is insufficient for detailed topography).
- Explore 3D landslide identification methods.
- Investigate the impact of different landslide types and potential errors in LULC data.
Summary updated