skill-assessment

Flood skill assessment

Framework to assess if and which meterological data (river discharge, rainfall, etc.) is of sufficient quality to be included in the trigger model. The assessment is based on how well a statistical model which uses metereological data can predict past floods. The model performance is measured by several indicators: False Alarm Rate (FAR), Probability Of Detection (POD), Critical Success Index (CSI).

Current version uses historical river discharge (source: Glofas) and rainfall data (source: CHIRPS).

Directory Structure

• scripts model and visualization scripts
• africa global input data for all Africa (Glofas virtual station, Hydroshed, etc.)
• uganda, kenya... input and output data per country

Setup

Requirements:

• Python 3.7.4

to install necessary modules, execute

pip install -r requirements.txt

Model

What does it do?

1. extract discharge data from Glofas

2. extract CHIRPS rainfall data from Google Earth Engine

3. train and test a model to predict floods

4. save model performance in a .CSV

How do I execute it?

to run the model, execute

python scripts/V12_glofas_analysis.py

V12_glofas_analysis.py accepts the command line arguments described below,

usage: V12_glofas_analysis.py [-h] [country] [ct_code] [model] [loss]

positional arguments:
  country     [Uganda]
  ct_code     [uga]
  model       [bdt_discharge_rainfall]
  loss        [far]

optional arguments:
  -h, --help  show this help message and exit

Which models are implemented?

• quantile_discharge based on thresholds with quantiles, using only glofas discharge data (best GloFAS station and threshold is computed per district)
• bdt_discharge based on decision trees, using only glofas discharge data
• bdt_discharge_rainfall based on decision trees, using glofas discharge data and rainfall

Visualization

How do I visualize model performance?

to visualize performance, execute

python scripts/IBF_flood_model_performance_visual.py

this will create maps of the performance or the model per district, by plotting FAR, POD, POFD, CSI and the number of available events per district.