TITLE: Near-real time landslide predictions in Colombia. Results of a case study in the Cauca, region.
TEAMS: Piquant rainbow lobster & Geoscripters 2023

WHY? - The aim or question of the project

The aim of our project is to use landslide point data of Colombia and relate various factors which may potentially contribute to increased landslide risk, such as: slope, precipitation, vegetation before the event and fault proximity. We would use this relationship to identify areas at risk of landslides in the Cauca region of Colombia in near-real time.

WHAT? - The datasets & metadata (author, date, extent, resolution)

Datasets	Purpose	Link	Author	Date	Extent	Resolution	Size
Landslide event point with coordinates and date of occurrence	Training points	https://simma.sgc.gov.co/#/	Colombian Geologic Service	2017 - 2022	Colombia		4.5MB
No Landslide points	Training points	Randomly selected in areas without landslides		2017 - 2022	Colombia
Slope	Predictor	https://developers.google.com/earth-engine/datasets/catalog/USGS_SRTMGL1_003	SRTM - NASA	2000	Colombia	30 m	GEE
Near real time precipitation	Predictor	https://developers.google.com/earth-engine/datasets/catalog/NASA_GPM_L3_IMERG_V06	NASA GES DISC	2000-2023	Colombia	10km	GEE
NDVI	Predictor	https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR	ESA- Sentinel 2	2016 - 2023	Colombia	10 m	GEE
Faults proximity	Predictor	https://datos-sgcolombiano.opendata.arcgis.com/datasets/e03339c845d24e7baceb6d67397a23b3_0.geojson?outSR=%7B%22latestWkid%22%3A3857%2C%22wkid%22%3A102100%7D	Colombian geologic service	2020	Colombia		4.5 MB

HOW? - Methods and potential results

Step by step

1. Retrieve landslide information. Where have they happened and when? (DONE for all Colombia from 2017 to 2022) [Excel file to geojson]

2. Create synthetic data of no landslides for areas where there have not been landslides in our time of study. (DONE - R code in repository)

3. Retrieve GEE information for both landslide and no landslide datasets (DONE)

   1. This is the process that takes the longest because it calculates NDVI, slope and accumulated precipitation of an area of interest around the point provided.
   2. The output here is a training dataset with 4 columns:
      1. Landslide_dummy: 0 or 1 if the event happened or not.
      2. NDVI: A value of NDVI from the week before the event occurred or not.
      3. Slope: A mean value of slope in the point of interest (POI).
      4. Precipitation: An accumulated value of precipitation for the week before the event.

4. Retrieve additional predictors, for example distance to faults. (R code to get faults in repo, but missing distance info and assign it to training dataset)

5. Train logistic regression model with training dataset (y = Landslide_dummy, X = all the predictors)

   1. Use of sci-kit learn package in Python to fit logistic regression.

6. Calculate odds ratio and present them as a result.

7. Apply the model for an area of interest. For this we would need:

   1. One cloud free (as free as we can find) sentinel 2 image to calculate the NDVI.
   2. Slope SRTM data
   3. Precipitation values (data from GPM) (Weekly accumulated)
   4. Fault proximity raster

8. Create a dynamic map of landslide prone areas in the area at the moment.

9. If we have time create an app to allow the user to decide an area where to vsualize the landslide prone zones.

Potential results

• Quantification of the effect of a change in predictors (increase or decrease in slope, precipitation, vegetation) in the odds of having a landslide.
• Near-real time probability map of landslide hazard in study area. "# GEE_data_retrieval_LS_Colombia"