The identification of areas that are prone to landslides is essential in mitigating associated risks. This is usually achieved using landslide susceptibility models, which estimate landslide likelihood given local terrain conditions and the location of known past events. Detailed databases covering different conditioning factors are paramount in producing reliable susceptibility maps. However, thematic data from developing countries are scarce. As a result, susceptibility models often rely on morphometric parameters that are derived from widely-available digital elevation models. In most cases, simple parameters, such as slope, aspect, and curvature, computed using a moving window of 3 × 3 pixels, are used. Recently, the use of window-based morphometric indices as an additional input has increased. These rely on a user-defined observation window size. In this contribution, we examine the influence of observation window size when using window-based morphometric indices as core predictive variables for landslide susceptibility assessment. We computed a variety of models that include morphometric indices that are calculated with different window sizes, and compared the predictive capabilities and reliability of the resulting predictions. All of the models are based on the random forest algorithm. The results improved significantly when each window-based morphometric index was calculated with a different and meaningful observation window (AUC-ROC of 0.89 and AUC-PR of 0.87). The sensitivity analysis highlights both the highly-informative observation windows and the impact of their selection on the model performance. We also stress the importance of evaluating landslide susceptibility results while using different adapted metrics for predictive performance and reliability.
[1]
Haibo He,et al.
Learning from Imbalanced Data
,
2009,
IEEE Transactions on Knowledge and Data Engineering.
[2]
Gaël Varoquaux,et al.
Scikit-learn: Machine Learning in Python
,
2011,
J. Mach. Learn. Res..
[3]
R. D. Hobson.
Surface roughness in topography: quantitative approach
,
2019,
Spatial Analysis in Geomorphology.
[4]
Andrew D. Weiss.
Topographic position and landforms analysis
,
2001
.
[5]
Yu Wang,et al.
A Debiased MDI Feature Importance Measure for Random Forests
,
2019,
NeurIPS.
[6]
Yoshua Bengio,et al.
Random Search for Hyper-Parameter Optimization
,
2012,
J. Mach. Learn. Res..