IRONHACK FINAL PROJECT: Predicting wildfires in California, two step approach

What is the project all about? The danger of wildfires on the West Coast, especially in California is increasing. Facing the multiple dimensions of climate change, one may ask: Is fire season now a year-round reality? To answer this question, I decided to work on a model predicting the amount and size of wildfires in California in a two step approach.

Please find the description of each Jupyter Notebook below:

(1) DATA MINING | NOAA API Scrapping data.ipynb: Scrapping weather data from NOAA API

• Setting up NOAA API for scrapping weather data of California, United States of America.
• Scrapping data from NOAA API was done from 1975 (for wind speed and evaporation starting 1985) to 2020.
• Please find documentation for NOAA API here: https://www.ncdc.noaa.gov/cdo-web/webservices/v2. Data is returned as json-format. The request for a chosen climate data per time period is automated and stored as pd.DataFrame for further use by each available weather station (4427 unique station IDs).

*To use the NOAA API you have to register to get an own token.

(2.1) DATA PROCESSING | UBER H3 for weather station & fires.ipynb: Mapping weather stations & wildfires with hexagons

• The H3 libary is used for indexing locations in California. Please find documentation for H3: Uber’s Hexagonal Hierarchical Spatial Index here: https://eng.uber.com/h3/.

• Weather stations are clustered per hexagon. For this model the resolution level 4 is chosen. Please find the created interactive map for available weather stations here: https://github.com/caaarov/finalproject_wildfires_CA/blob/master/files/map_weather%20stations%20CA.html

• Recorded wildfires between 1992-2015 in California are clustered per hexagon. For this model the resolution level 4 and level 5 are chosen. Please find the created interactive map on level 5 for wildfires here: https://github.com/caaarov/finalproject_wildfires_CA/blob/master/files/map_wildfires_level%205_CA.html

(2.2) DATA PROCESSING | Time Series Analysis with PROPHET.ipynb: Time Series Analysis with PROPHET is documented for the climate data of Temperature | Precipitation | Wind Speed | Evaporation.

• The Time Series Analysis with PROPHET was done for the next five years (2021 - 2026).
• Please find the documentation for PROPHET here: https://facebook.github.io/prophet/docs/quick_start.html#python-api. The forecast for 2021-2026 is automated and stored as pd.DataFrame for further use.

(3.1) MODELING STEP 1 | Probability of wildfires.ipynb: In this jupyter notebook the deployment of STEP 1 for the model of wildfires prediction is documented.

• Data Exploration: Between 1992 – 2015 more than 1.8 million wild fires were recorded. Find e.g. details about fire size and cause in section 1.
• Feature Engineering: Urban areas are included, climate data of the last quarter & year is computed.
• Balancing of Data: Manual balancing between fire size classes to reduce the false negative predictions of the model (increasing the recall score).
• Automated Feature Selection: Tuning down the number of features, weather variables were identified as main features.
• Grid Search for AdaBoostClassifier: Tuning the hyper-parameters of AdaBoostClassifier.

The accuracy of the model in the TEST set is: 0.97

The kappa of the model in the TEST set is: 0.894

For 2020, the model predicts a record number of fires for the San Francisco area. By the end of 2020, 9,639 fires had burned 4,397,809 acres – more than 4% of California's acres of land. For 2020, the model predicts a record number of fires for the area around San Francisco.

(3.2) MODELING STEP 1 | Predicting wildfire size.ipynb: In this jupyter notebook the deployment of STEP 2 for the model of wildfires prediction is documented.

** IN PROGRESS **

• Number of fires between 4 – 20 km2 is increasing in California.
• Automated Feature Selection: Tuning down the number of features.
• For the prediciton of the wildfire size the imbalance of the size of recorded wildfires is challenging.

718101 A (- 0.25 acres)

606186 B (0.26 - 9.9 acres)

64670 C (10 - 99.9 acres)

13231 D (100 - 299 acres)

6465 E (300 - 999 acres)

4400 F (1000 - 4999 acres)

2696 G (5000 acres)

• Model is still under evaluation and not sophisticated enough for a reliable prediction of fire size.

(4.1) STORY TELLING | Data base SAN FRANCICSO.ipynb

In this Jupyter Notebook the data generation for the final project presentation is documented.

(4.2) STORY TELLING | Use Case SAN FRANCISCO.ipynb

In this Jupyter Notebook the data generation for the final project presentation is documented.

Predicted wildfires between 2016-2020 for ring 2 nearst neighbours of hexagon of San Francisco are summed. For this model the resolution level 4 is chosen. Please find the created interactive map on level 4 for wildfires in the Bay Area here: https://github.com/caaarov/finalproject_wildfires_CA/blob/master/files/map_2021-predictions_San%20Francisco%20Area%20.html

Find the PDF of my presentation on GitHub: https://github.com/caaarov/finalproject_wildfires_CA/blob/master/Wildfires%20in%20California%20-%20A%20Machine%20Learning%20Model.pdf.

IDEAS FOR EXTENDING THE MODEL:

• Setting up a live connection to NOAA API to keep the model with live climate & weather data up to date.
• Adding more feature with information about the ecosystem and biodiversity of the hexagons.
• Labeling hexagons with high percentage of land covered by forest areas.
• Using level 5 or 6 for H3 indexing for higher granularity.
• Using NN to find "hidden features" (like regulations of the state in specific time periods to lower the wildfire risk).

...