<p>Estimation of landslide susceptibility in mountainous areas is a prerequisite for risk assessment and contingency planning. The susceptibility to landslide is modelled based on thematic layers of information such as geomorphology, hydrology, or geology, where detailed characteristics of the area are depicted. The growing use of machine learning techniques to identify complex relationships among a high number of variables decreased the time required to distinguish areas prone to landslides and increased the reliability of the results. However, numerous countries lack detailed thematic databases to feed in the models. As a consequence, susceptibility assessment often relies heavily on geomorphic parameters derived from Digital Elevation Models. Simple parameters such as slope, aspect and curvature, calculated under a moving window of 3x3-pixels are mostly used. Furthermore, advanced morphometric indices such as topographic position index or surface roughness are increasingly used as additional input parameters. These indices are computed under a bigger window of observation usually defined by the researcher and the goal of the study. While these indices proved to be useful in capturing the overall morphology of an entire slope profile or regional processes, little is known on how the selection of the moving window size is relevant and affects the output landslide susceptibility model.&#160;</p><p>In order to address this question, we analysed how the predicting capabilities and reliability of landslide susceptibility models were impacted by the morphometric indices and their window of observation. For this purpose, we estimate the landslide susceptibility of an area located in Tajikistan (SW Tien Shan) using a Random Forest algorithm and different input datasets. Predicting factors include commonly used 3x3-pixel morphometrics, environmental, geological and climatic variables as well as advanced morphometric indices to be tested (surface roughness, local relief, topographic position index, elevation relief ratio and surface index). Two approaches were selected to address the moving window size. First, we chose a common window of observation for all the morphometric indices based on the study area valley&#8217;s characteristics. Second, we defined an optimal moving window(s) for each morphometric index based on the importance ranking of models that include moving windows from a range of 300 to 15000 m for each index. A total of 20 models were iteratively created, started by including all the moving windows from all the indices. Predicting capabilities were evaluated by the receiver operator curve (ROC) and Precision-Recall (PR). Additionally, a measure of reliability is proposed using the standard deviation of 50 iterations. The selection of different moving windows using the feature importance resulted in better-predicting capabilities models than assigning an optimal for all. On the other hand, using a single different moving window per morphometric index (eg. most important ranked by random forest) decreases the evaluating metrics (a drop of PR from 0.88 to 0.85). Landslide susceptibility models can thus be improved by selecting a variety of meaningful (physically and methodological) windows of observation for each morphometric index. A 3x3-pixel moving window is not recommended because it is too small to capture the morphometric signature of landslides.&#160;</p>
High-resolution snow depth prediction using Random Forest algorithm with topographic parameters and an ecosystem map: a case study in the Greiner Watershed, Nunavut
