Explainable artificial intelligence in geoscience: a glimpse into the future of landslide susceptibility modeling

Dahal, Ashok; Lombardo, Luigi

doi:10.1002/essoar.10512130.1

Cited by 3 publications

(3 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…So far, this level of model explainability was already presented in three recent articles (Collini et al, 2022;Zhang et al, 2023;Dahal and Lombardo, 2022). However, what they all missed is translating the information offered by the SHAP values across the geographic space, which is what we will present in the next section.…”

Section: Local Interpretationmentioning

confidence: 76%

On the use of explainable AI for susceptibility modeling: examining the spatial pattern of SHAP values

Wang¹,

Zhang²,

Dahal³

et al. 2023

Preprint

View full text Add to dashboard Cite

Hydro-morphological processes (HMP, any natural phenomenon contained within the spectrum defined between debris flows and flash floods) are globally occurring natural hazards which pose great threats to our society, leading to fatalities and economical losses. For this reason, understanding the dynamics behind HMPs is needed to aid in hazard and risk assessment. In this work, we take advantage of an explainable deep learning model to extract global and local interpretations of the HMP occurrences across the whole Chinese territory. We use a neural network architecture and interpret the model results through the spatial pattern of SHAP values.In doing so, we can understand the model prediction on a hierarchical basis, looking at how the predictor set controls the overall susceptibility as well as doing the same at the level of the single mapping unit.Traditional statistical tools usually lead to a clear interpretation at the expense of large performance, which is otherwise reached via machine/deep learning solutions, though at the expense of interpretation. Explainable AI is the key to combine both strengths and in this work, we explore this combination in the context of HMP susceptibility modeling. Specifically, we demonstrate the extent to which one can enter a new level of data-driven interpretation, supporting the decision-making process behind disaster risk mitigation and prevention actions.

show abstract

Section: Local Interpretationmentioning

confidence: 76%

On the use of explainable AI for susceptibility modeling: examining the spatial pattern of SHAP values

Wang¹,

Zhang²,

Dahal³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Four susceptibility models were generated for each of the modelling techniques, using different proportions of the inventory: (a) 25%, (b) 50%, (c) 75% and (d) 100%. The susceptibility models of the LMT, J48, RF and NBTree algorithms are shown in Figures 6, 7, 8 and 9, respectively, as presented by Dahal and Lombardo (2023). From all the figures, it is evident that a similarity exists among the maps.…”

Section: Resultsmentioning

confidence: 99%

Landslide risk assessment and management using hybrid machine learning‐based empirical models

Yang,

Jiang,

Arabameri

et al. 2023

Geological Journal

View full text Add to dashboard Cite

Landslides are a prevalent geologic phenomenon that substantially threatens human life and infrastructure, resulting in considerable loss and destruction. The practice of landslide susceptibility mapping is crucial for the mitigation of risks connected with this natural disaster. This work aims at investigating the influence of varying sample sizes on the precision of landslide susceptibility modelling using a case study conducted in the Alamout basin, Iran. The researchers used a machine learning methodology based on tree algorithms to construct a model for predicting the likelihood of landslides. Additionally, they adopted a multi‐scenario strategy to address the inherent uncertainty associated with the input data. The integration of the naive Bayes tree (NBTree), random forest (RF), logistic model tree (LMT) and J48 algorithms was performed. The modelling process included using 20 predictive parameters across four distinct scenarios. Four models, labelled S1, S2, S3 and S4, were used in this study. These models utilized 25%, 50%, 75% and 100% of the available inventory data. The research presented in this study is distinguished by using a tree‐based methodology for landslide susceptibility modelling and incorporating a multi‐scenario strategy to address the inherent uncertainty associated with the input data. The findings indicated that the augmentation of the sample size improved the precision of the models. The efficacy of using a multi‐scenario strategy in enhancing the dependability of the model is also underscored. Among the 20 input elements used in the modelling process, it was seen that slope angle accounted for the highest relative significance, constituting 25.60% of the overall influence. Following more closely, distance to fault contributed significantly, with a relative importance of 23.40%. Additionally, rainfall and elevation exhibited notable contributions, with relative volumes of 7.91% and 5.50%, respectively. All four landslide models showed adequate learning and forecasting ability throughout the training and testing phases. During the testing phase, the true skill score (TSS) values exhibited a range of 0.631–0.804, while the area under the receiver operating characteristic curve values showed a range of 0.745–0.921. The susceptibility maps indicated that a significant portion of the region exhibits moderate to very high susceptibility zones, with the northern and eastern sectors displaying greater landslide values than the western region. The model's performance showed improvement from S1 to S4 in both the training and testing phases. The performance of the models exhibited the following trend: in scenario 1, the RF model outperformed the J48, LMT and NBTree models; in scenario 2, the RF model surpassed the NBTree and LMT models, while being on par with the J48 model; in scenarios 3 and 4, the RF model showed superior performance compared to the NBTree, J48 and LMT models. Therefore, the RF model proved to be the most effective among the models evaluated. The findings derived from this research have the potential to serve as valuable references for the purposes of land‐use planning and catastrophe risk management.

show abstract

“…From the slope unit map, all the areas below the 10 • slope were removed to exclude flat areas from our analysis. Details on the slope unit generation are provided in Dahal and Lombardo (2022).…”

Section: Observations and Spatial Domainmentioning

confidence: 99%

From ground motion simulations to landslide occurrence prediction

Dahal¹,

Cruz²,

Tanyaş³

et al. 2023

Preprint

View full text Add to dashboard Cite

Ground motion simulations solve wave equations in space and time, thus producing detailed estimates of the shaking time series. This is essentially uncharted territory for geomorphologists, for we have yet to understand which ground motion (synthetic or not) parameter, or combination of parameters, is more suitable to explain the coseismic landslide distribution. To address this gap, we developed a method to select the best ground motion simulation using a combination of Synthetic Aperture Radar Interferometry (InSAR) and strong motion data. Upon selecting the best simulation, we further developed a method to extract a suite of intensity parameters, which we used to both bivariately and multivariately analyse coseismic landslide occurrences taking the Gorkha earthquake as a reference. Our results show that beyond the virtually unanimous use of peak ground acceleration, velocity, or displacement in the literature, different shaking parameters could play a more relevant role in landslide occurence. These parameters are not necessarily linked to the peak values but mostly linked to the actual displacement, velocity, frequency content and shaking duration, elements too often neglected in geomorphological analyses. This in turn implies that we have yet to fully acknowledge the complexity of the interactions between full waveforms and hillslope responses.

show abstract

Explainable artificial intelligence in geoscience: a glimpse into the future of landslide susceptibility modeling

Cited by 3 publications

References 46 publications

On the use of explainable AI for susceptibility modeling: examining the spatial pattern of SHAP values

On the use of explainable AI for susceptibility modeling: examining the spatial pattern of SHAP values

Landslide risk assessment and management using hybrid machine learning‐based empirical models

From ground motion simulations to landslide occurrence prediction

Contact Info

Product

Resources

About