GEO-AI Challenge for Landslide Susceptibility Mapping by AI for Good ‐ International Telecommunication Union
This project achieved the 1st place on Zindi's GEO-AI Challenge for Landslide Susceptibility Mapping competition!!!
This repository contains the code used for my submission for Zindi's GEO-AI Challenge for Landslide Susceptibility Mapping by ITU competition. The goal of the challenge was to develop an accurate and cost-effective model for landslide susceptibility mapping in the Valtellina Valley region in Italy using machine learning techniques.
You can watch my presentation on the challenge finale, where I explain my approach. Also later in the video you can see my 1st place certificate!
My submission for the competition used the only the official datasets provided by the competition hosts. In order to engineer features I created a circular buffer zone for each data point with the following radii: 100 m, 500 m, 1 km, 2.5 km, gathering information and relevant characteristics present in these areas. This is done because given that we want to model landslide susceptibility, the information present directly within the geometry that we're considering might not be enough. For instance, the fact that there's a river 50 m away from the geometry we're analyzing might make it more susceptible to a landslide, even though the river isn't crossing directly thorough the geometry. After acquiring and engineering the features, I used a LightGBM model to fit the data, performing a hyperparameter search with Optuna with 5-fold cross-validation to achieve the best performance. After 100 optimization trials, the model achieved an accuracy of 96.82% on the 5-fold cross-validation, leading to a score of 94.67% on the final competition leaderboard.
main.ipynb: Jupyter notebook file containing my complete analysis, thoroughly explained. I recommend using the following nbviewer link to visualize it online.GEO-AI Challenge for Landslide Susceptibility Mapping report.pdf: an executive summary explaining the development of my features and model.colorful_100.qgz: final QGIS susceptibility map, where 1 are areas with very low probability of landslides and 5 are areas with very high probability.map.png: picture from the final susceptibility map. Green are areas with very low probability and red are areas with very high probability. The points indicate the location of the test set data.data: directory containing all of the data used.datasets: directory where the data provided by the competition hosts should go.predictions.csv: final predictions used for the competition.*(other files): intermediate files that can be useful to reproduce the code.