My solution to the Inundata: Mapping Floods in South Africa challenge by Zindi. The goal of the challenge was to predict flash floods from CHIRPS precipitation data and Sentinel 2 images.
# Inundata-Mapping-Floods-in-South-Africa
My solution to the Inundata: Mapping Floods in South Africa challenge by Zindi. The goal of the challenge was to predict flash floods from CHIRPS time series precipitation data and Sentinel 2 images.
Each time series has 760 steps and at most one flood event. As per the event organizers, probabilities cannot be thresholded or rounded.
The Feature Engineering notebook has a small EDA and creates the necessary features for prediction. Mainly lags, leads, and different rolling statistics. Also, for the images, Vegetation Indexes and Gray-Level Co-Occurrence Matrix features are calculated.
The Modelling notebook has the 10 fold cross validation, using LightGBM for a model that predicts whether there's a flash flood event in a time series or not, and also for a model that predicts WHEN a flash flood might occur.
Train and test data are provided, while images have to be downloaded from Zindi's website. Also, the submission notebook is not included since the challenge has already closed.