Periodic Reporting for period 1 - HYBLAND (Development of a hybrid methodology for the susceptibility and hazard analysis of landslides)
Période du rapport: 2022-04-18 au 2024-04-17
Predicting the activation and consequences of geohazards, such as landslides, is challenging due to large uncertainties in geological and geotechnical conditions and human activities. The HYBLAND project aims to develop a novel, reliable, and comprehensive method for estimating landslide susceptibility and hazard. The study's objectives include: a) investigating the correlation between geotechnical factors and landslide occurrence, b) formulating an effective multimodal deterministic approach for landslide susceptibility assessment, c) developing a landslide susceptibility and hazard assessment method using Machine Learning (ML), and d) combining these methods to create a highly performing approach. This new method manages uncertainties effectively, providing reliable results even in areas with scarce landslide records.
The test area for developing this approach is southwestern Cyprus, characterized by a high density of landslides due to its geology and seismicity. Once validated, the method and resulting maps can be applied to other European regions with similar issues. The research outcomes will assist decision-makers in reliably protecting the built environment from landslide hazards, ensuring sustainable development.
The test area for developing this approach is southwestern Cyprus, characterized by a high density of landslides due to its geology and seismicity. Once validated, the method and resulting maps can be applied to other European regions with similar issues. The research outcomes will assist decision-makers in reliably protecting the built environment from landslide hazards, ensuring sustainable development.
Initially, all necessary data for the testbed area were collected and digitized to create thematic maps with a 15m×15m grid. A geotechnical database was constructed using data and field reports from the Geological Survey Department of Cyprus, processing over 700 geotechnical boreholes. Fieldwork campaigns were conducted in areas lacking geotechnical data, where field measurements were taken and samples collected for laboratory testing.
To enhance the landslide inventory map, InSAR analysis for 2017-2021 was conducted using the Small Baseline Subset (SBAS) technique. In addition, an innovative method for landslide detection, leveraging a boosted cascade of simple geomorphological features within a ML framework, was developed to efficiently identify areas with existing landslides (active, dormant, and relict). Findings indicate that high-resolution DEMs and smaller cell sizes are most effective for identifying relict landslides with subtle features, albeit at high computational costs. For active landslides with steeper geomorphologies, a coarser cell size combined with moderate-resolution DEMs achieves an 80% success rate.
In the next project phase, a detailed assessment of the influence of seventeen causal factors was conducted. These factors were digitized into thematic layers in ArcGIS with a 15m × 15m grid size. This process involved two steps. First, the influence degree of these factors was evaluated across the entire research area using the mean decrease accuracy value in the Random Forest (RF). Second, a smaller area of 552.27 km² was selected to study the main geotechnical and geomorphological factors affecting slope stability. Statistical correlations using the Frequency Ratio method and advanced ML techniques (RF and Shapley Additive Explanations) indicate that geotechnical factors are more critical than geomorphological factors in landslide activation.
The development of landslide susceptibility maps using ML methods involved several subroutines, including the creation of landslide and no-landslide inventory databases, the preparation of training and testing datasets, and the classification of landslide-triggering factors into susceptibility classes. Statistical processing to construct the susceptibility maps was performed in R using three methodologies: Logistic Regression (LR), RF, and Extreme Gradient Boosting (XGB). Each model's validation and accuracy were assessed using methods such as Receiver Operating Characteristics and the Wilcoxon signed-rank test. The LR model achieved a testing accuracy of 92.4%, RF 94.3%, and XGB 93.2%. Regarding predictive capability, LR, XGB, and RF identified 87.0%, 96.3%, and 98.7% of recorded landslides as high to very high susceptibility areas, respectively. Thus, the RF demonstrated the best performance among the three.
The project's deterministic approach uses rough estimates of the factor of safety (FS) calculated at each grid point using closed-form equations. At each grid point, a cross-section of the ground is created along the direction of the slope inclination. and the ground surface profile is approximated by an equivalent uniform slope having constant inclination. Two failure modes are considered for each grid point: rotational and planar, with planar failure assumed to occur as slippage on bedding planes.
Upon concluding the project, a new hybrid methodology was developed by integrating probabilistic RF and deterministic methods. This involved using the deterministic FS map as a thematic layer within the probabilistic approach. To evaluate the hybrid method's performance, three separate analyses were performed:
1.Probabilistic analysis using 13 causal factors commonly identified in landslide susceptibility literature.
2.Addition of four geotechnical factors (GSI, plasticity index, clay content, dip direction difference) to the probabilistic method, alongside the 13 factors from the first set.
3.Probabilistic analysis incorporating the 13 common causal factors and the FS from the deterministic method.
In the third set of analyses, it was found that the FS ranks in second place after the lithology in terms of importance, with the latter having significantly reduced importance compared to the 1st and 2nd sets of analyses. Crucially, it is evident that using deterministic approach results as the thematic layer in the probabilistic approach provides optimal performance
Outreach activities were conducted according to the predetermined plan through publications and workshops.
To enhance the landslide inventory map, InSAR analysis for 2017-2021 was conducted using the Small Baseline Subset (SBAS) technique. In addition, an innovative method for landslide detection, leveraging a boosted cascade of simple geomorphological features within a ML framework, was developed to efficiently identify areas with existing landslides (active, dormant, and relict). Findings indicate that high-resolution DEMs and smaller cell sizes are most effective for identifying relict landslides with subtle features, albeit at high computational costs. For active landslides with steeper geomorphologies, a coarser cell size combined with moderate-resolution DEMs achieves an 80% success rate.
In the next project phase, a detailed assessment of the influence of seventeen causal factors was conducted. These factors were digitized into thematic layers in ArcGIS with a 15m × 15m grid size. This process involved two steps. First, the influence degree of these factors was evaluated across the entire research area using the mean decrease accuracy value in the Random Forest (RF). Second, a smaller area of 552.27 km² was selected to study the main geotechnical and geomorphological factors affecting slope stability. Statistical correlations using the Frequency Ratio method and advanced ML techniques (RF and Shapley Additive Explanations) indicate that geotechnical factors are more critical than geomorphological factors in landslide activation.
The development of landslide susceptibility maps using ML methods involved several subroutines, including the creation of landslide and no-landslide inventory databases, the preparation of training and testing datasets, and the classification of landslide-triggering factors into susceptibility classes. Statistical processing to construct the susceptibility maps was performed in R using three methodologies: Logistic Regression (LR), RF, and Extreme Gradient Boosting (XGB). Each model's validation and accuracy were assessed using methods such as Receiver Operating Characteristics and the Wilcoxon signed-rank test. The LR model achieved a testing accuracy of 92.4%, RF 94.3%, and XGB 93.2%. Regarding predictive capability, LR, XGB, and RF identified 87.0%, 96.3%, and 98.7% of recorded landslides as high to very high susceptibility areas, respectively. Thus, the RF demonstrated the best performance among the three.
The project's deterministic approach uses rough estimates of the factor of safety (FS) calculated at each grid point using closed-form equations. At each grid point, a cross-section of the ground is created along the direction of the slope inclination. and the ground surface profile is approximated by an equivalent uniform slope having constant inclination. Two failure modes are considered for each grid point: rotational and planar, with planar failure assumed to occur as slippage on bedding planes.
Upon concluding the project, a new hybrid methodology was developed by integrating probabilistic RF and deterministic methods. This involved using the deterministic FS map as a thematic layer within the probabilistic approach. To evaluate the hybrid method's performance, three separate analyses were performed:
1.Probabilistic analysis using 13 causal factors commonly identified in landslide susceptibility literature.
2.Addition of four geotechnical factors (GSI, plasticity index, clay content, dip direction difference) to the probabilistic method, alongside the 13 factors from the first set.
3.Probabilistic analysis incorporating the 13 common causal factors and the FS from the deterministic method.
In the third set of analyses, it was found that the FS ranks in second place after the lithology in terms of importance, with the latter having significantly reduced importance compared to the 1st and 2nd sets of analyses. Crucially, it is evident that using deterministic approach results as the thematic layer in the probabilistic approach provides optimal performance
Outreach activities were conducted according to the predetermined plan through publications and workshops.
The main novel aspect can be summarized as followed:
• The two basic methodologies, were combined for the first time to produce a hybrid approach. The synergy between the two enhances their merits and alleviates their shortcomings.
• For the first time, geotechnical factors (GSI, unconfined compressive strength, plasticity index, clay content) were used as layers in stochastic analysis for constructing landslide susceptibility maps.
• Regarding the automatic landslide detection procedure, the approach formulated in the context of the project trains a ML algorithm on the basis of the hillshade of the study area. The proposed approach was successful in identifying even relict landslides despite their relatively subtle geomorphological features.
• Novel features of the developed deterministic approach include the consideration of the effect of slope height, the reduced ground strength in areas with existing landslides, and the effect of pore water pressures through the distance from the hydrographic network.
The HYBLAND results are crucial for strategic sustainable development planning. Specifically, this research aids local authorities and decision-makers in implementing safer measures to protect urban areas from landslides, such as avoiding high-risk zones, dewatering, and soil erosion protection. The Pissouri landslide incident, where inadequate urban planning led to significant financial losses, underscores the urgent need for this research. Additionally, the generated maps will guide private stakeholders, including construction companies and real estate developers, in selecting lower-risk development sites, thereby enhancing long-term profits by reducing future financial liabilities.
• The two basic methodologies, were combined for the first time to produce a hybrid approach. The synergy between the two enhances their merits and alleviates their shortcomings.
• For the first time, geotechnical factors (GSI, unconfined compressive strength, plasticity index, clay content) were used as layers in stochastic analysis for constructing landslide susceptibility maps.
• Regarding the automatic landslide detection procedure, the approach formulated in the context of the project trains a ML algorithm on the basis of the hillshade of the study area. The proposed approach was successful in identifying even relict landslides despite their relatively subtle geomorphological features.
• Novel features of the developed deterministic approach include the consideration of the effect of slope height, the reduced ground strength in areas with existing landslides, and the effect of pore water pressures through the distance from the hydrographic network.
The HYBLAND results are crucial for strategic sustainable development planning. Specifically, this research aids local authorities and decision-makers in implementing safer measures to protect urban areas from landslides, such as avoiding high-risk zones, dewatering, and soil erosion protection. The Pissouri landslide incident, where inadequate urban planning led to significant financial losses, underscores the urgent need for this research. Additionally, the generated maps will guide private stakeholders, including construction companies and real estate developers, in selecting lower-risk development sites, thereby enhancing long-term profits by reducing future financial liabilities.