FireWatch

Policy-based wildfire prevention. $10k pre-seed from Renewal funds, InvestEco and Vancity.

Currently, we are in the process of building a predictive model using open-source data in California, combining multistage ML models. We hope to incorporate direct factors along with human factors, such as campgrounds, to effectively predict and manage wildfire resources. Version 1.0

More links

• YIC Video Submission
• Peel Media Article

Initial fire plotting

Here is the initial plotting of wildfires in California, alongside a Gaussian distribution to create a probability, P, where 0 < P < 1.

Factors mapping

Right now, we are building the model using 2 factors: 1 direct (temperature) and 1 neighbouring (campgrounds).

To obtain temperature data, we check the center of each grid cell and call on the weather API.

temperature_data = []
fishnet.to_crs(epsg=4326, inplace=True)
...
for cell_polygon in fishnet['geometry']:
    centroid = cell_polygon.centroid
    latitude, longitude = centroid.y, centroid.x  # Note that y is latitude and x is longitude
    new_temp = temp(latitude, longitude)
    temperature_data.append(new_temp.temperature[1])
...

Normalizing and assigning a color via cmap.

For the campgrounds factor, the campground is plotted, and then the nearest distance is recorded and mapped via a cmap.

for cell_polygon in fishnet['geometry']:
    centroid_x, centroid_y = cell_polygon.centroid.x, cell_polygon.centroid.y

    intersecting_campgrounds = campgrounds_in_california[campgrounds_in_california.intersects(cell_polygon)]
    if not intersecting_campgrounds.empty:
        campground_coordinates.append([centroid_x, centroid_y])

for cell_polygon in grid_polygons:
    centroid_x, centroid_y = cell_polygon.centroid.x, cell_polygon.centroid.y
    ...
    distance = max_number
    for campground in campground_coordinates:
        if distance_between([centroid_x, centroid_y], campground) < distance:
            distance = distance_between([centroid_x, centroid_y], campground)
    fishnet_distance.append(distance)

Local Moran

Applying a local Moran helps to reduce noise within the dataset.

w = libpysal.weights.Queen.from_dataframe(fishnet)
moran_loc_wildfire = Moran_Local(wildfire_probability, w, permutations=999)
fishnet['local_moran_wildfire'] = moran_loc_wildfire.Is

Here is the Moran overlayed with 2023 wildfires after training.

Training

Using the kNN distance, logistic regression on temperature combined with the local Moran, a heatmap can be generated overlaying 2023 wildfires (not trained on) - achieving 74% accuracy on a 25000 square kilometer cell size.