GEO-AI Challenge for Landslide Susceptibility Mapping

Submission by Isaac Oluwafemi Ogunniyi (zindi username: isaacOluwafemiOg) for the ITU GeoAI Landslide Susceptibility Mapping challenge

This solution makes use of five files:

• 4 data files (all from the data provided as part of the competition):

  • 'Train.gpkg' which contains geometries classified as either susceptible or not.
  • 'geological_faults.gpkg' and 'land_use_land_cover.gpkg' for extracting additional features about any specific land geometry that will aid in the prediction of its susceptibility to landslides.
  • 'Test.gpkg' contains the data of locations for testing the performance of the model

• A Jupyter notebook named 'GeoLandslide.ipynb' containing the code for handling data, training model and predicting the landslide susceptibility of the test locations.

The environment

The Jupyter notebook runs in a python environment. For a Google Colab CPU runtime type, it lasts under 40 minutes. No paid subscriptions or resources were used

The Jupyter Notebook

• This notebook takes input of all 4 data files mentioned previously and outputs 'Submission.csv'.
• To run the notebook the 'data_path'' variable must contain the path to the directory containing all the data files mentioned previously.
• The notebook contains python code and the whole code runs on Google colab lasting for under 40 minutes.
• The General Overview of the Notebook is as follows:

1. Importing Relevant libraries geopandas (0.13.2), pandas (1.5.3), numpy (1.23.5), scikit-learn (1.2.2), imbalanced-learn (0.10.1)
2. Reading and Cleaning Data
3. Engineering Features The features fed into the model are 38 and fall into 3 categories:
   • Features from the train and test geometries which include the area ('area') of the geometry under consideration
   • Features from the land-use-land-cover categories which make up the geometry for which the susceptibility is being predicted. This includes total area of each of the land-use-land-cover category (eg: 'area_31' indicates the size of lands with land-use-land-cover code of 31 ('cod_31') that fall within the geometry under consideration)
   • Features from geological fault lines that lie within the geometry for which susceptibility is being predicted. This includes the cumulative length ('SHAPE_LEN') of the fault lines that falls within the geometry.
4. Developing Pipelines and training models
   • Pipeline involved scaling (standardScaler), selecting best features (SelectKBest) and then fitting a RandomForestClassifier model
   • Prior to fitting the model to the training data, balance was introduced to the number of landslide susceptible and landslide-proof observations by the help of the SMOTE technique.
   • The model's hyperparameters were tuned using the RandomizedsearchCV
5. Predicting and preparing submission
   • Predictions are made on the data imported from 'Test.gpkg' which has been transformed to include the same features as the train data.
   • The final predictions are exported to 'Submission.csv' in the same directory as the jupyter notebook.