Using Random Forest to predict avalanche runout

04.02.2021 - Håvard Toft Larsen

Table of contents

• Introduction
• Data preparation
• Random Forest

Introduction

Using the random forest algorithm, we developed four models for predicting alpha-angle runout from topographic parameters.

To redo our data preparation, the following data is needed:

Switzerland

• Avalanche dataset from Switzerland, contact Yves Bühler at SLF (https://www.slf.ch/en/index.html).
• SwissALTI3D_5m from SwissTopo for all of Switzerland (https://shop.swisstopo.admin.ch/en/products/height_models/alti3D)

Norway

• Avalanche dataset from Norway, contact Markus Eckerstorfer at NORCE (https://www.norceresearch.no/).
• National 10m DEM of Norway (https://hoydedata.no/LaserInnsyn/)

Data preparation

Switzerland

• Calculate the centerline of each avalanche using calculate_centerline.py (ESRI software is needed).
• Use ESRI ArcMap tool Generate Points Along Lines to generate a point every 5 meter of each avalanche centerline. Each point must include avalanche reference ID and elevation.
• Resample the data to 100 z-values for each avalanche using centerline_to_array.ipynb.

Calculate topograghic parameters

• y'' can be calculated using point_calculate_second_derivative.ipynb
• R, T and D can be calulcated using calculate_avalanche_confinement.py (ESRI software is needed).
• alpha, beta, theta can be calculated using array_calculate_alpha_beta.ipynb.
• path_type can be calculated using array_KMeans_clustering.ipynb.
• L_flow/linear, P, H, aspect, d_size, alt_min/max, are either given in the avalanche dataset or can be extracted using simple GIS tools.
• Hy'' and P/L_flow can be calculated from the topographic parameters above.

Norway

• Calculate potential release areas (PRA, https://github.com/jocha81/Avalanche-release)
• Use the TauDEM to model the alpha angle for each PRA (https://hydrology.usu.edu/taudem/taudem5/)
• Use calculate_percentiles.py to calculate the 5th percentile lowest alpha angle within the avalanche runout polygon.

Random Forest

• The whole model including training and target data is available. The model is available as a Jupyter Notebook or Google Colab: Thesis_main.ipynb, input data is main_data.csv