ClassiFIRE

Samuel Alter's BrainStation 2023 Data Science Capstone Project, Spring 2023

Motivation

• Wildfires, or the uncontrolled burning of vegetation, cause enormous damage and loss of life worldwide every year.¹
• 33 people died and >4.3 million acres burned in California in 2020 alone. This is equivalent to the total combined area of Puerto Rico and Rhode Island.
• Although they are a natural component of some forest ecosystems, wildfire incidents are projected to increase in our warming world.²
• While society transitions to a greener future, there is a clear need to predict where wildfires are likely to occur so that communities can protect themselves and evacuate when necessary.

Introduction

• For this project, I collected a combination of satellite imagery, topographic data (aspect, elevation, and slope), and historical fire boundaries to train models that would predict wildfire risk (Figure 1).
• I first used QGIS, an open-source mapping application, to setup the relationships between the datasets.
• Then used Python to process and model the data.
• I fed the processed data into machine learning and deep learning models to make my predictions.
• I chose to focus on the Santa Monica Mountains in Ventura and Los Angeles Counties, for its relatively manageable size, proximity to populated areas, and its east-west trend, which makes the topographic analysis easier.

Datasets Used

• Satellite imagery and topographic data from USGS’ EarthExplorer³
• Historical records of wildfire burn boundaries from the National Interagency Fire Center⁴
• File formats used: .tif and .jpg (imagery); .geojson and .csv (historical wildfire data).

Mapping, Data Cleaning, and Preprocessing

• In order to train a model that integrates spatial information with historical fire boundaries, I had to find a way to quantify the continuous spatial information. I settled on building an evenly-spaced points layer that would have the topographic and fire boundary data from under that point appended to the file (Figure 2).
• Since the satellite photos were from 2018, I removed all fire polygons from after 2018 and merged the rest to one shape.
• Using the point layer, an overlapping function could determine if a part of the landscape experienced a fire (“fire” area) or not (“nofire” area). The resultant file would serve as the “topographic” data in the project.
• For the geographic data, I focused on aspect (i.e. what direction the land is facing), elevation (meters above sea level), and slope (degrees). It is a straightforward operation to extract these data from an elevation raster and append them to the point layer.
• I tiled the satellite imagery into 128x128 pixels that would be fed into a Tensorflow image model, and saved them into two folders (i.e., fire and nofire).
• At the end of the data collection step, I had a table of geographic data and two image folders (fire areas and non-fire areas).

Modeling, Results, and Insights

• Since data is in two forms (point-based topographic information and satellite image tiles), I trained two models. After finding an optimal model, I created a metamodel, which used the predictions from the topographic and image analysis models as factors for a new logistic regression (Figure 3).

• Topographic Data
  • The areas affected by fire were predominantly located in mountainous regions with higher mean elevation compared to fire-free areas (Table 1).
  • Correlations were observed between different topographic factors, such as categorical aspect with continuous aspect, and categorical slope and elevation with continuous elevation and slope.
  • LogisticRegression was chosen as the modeling approach for classifying fire incidence based on topographic data, revealing the relative importance of different features.
  • Various models were tested, including naive_bayes, Bernoulli and Gauss, XGBClassifier, RandomForestClassifier, AdaBoost, and GradientBoostingClassifier, with GradientBoostingClassifier achieving the highest accuracy.
  • Grid search was employed to optimize the GradientBoostingClassifier model, but the best accuracy remained the same with specific parameter values.

• Satellite Imagery
  • BigEarthNet⁵ suggested the pre-trained VGG19 model to be sufficient, so I used that with Tensorflow-Keras and added a final dense layer with two output nodes to represent the fire/nofire categories required by this project.
  • With 20,025,410 total parameters, I deemed the model more than sufficient for the project’s needs.
  • After training, the model achieved an accuracy of 95.6% on classifying all the images.
  • Figure 4 shows typical images in the set as well as the locations of the two areas (fire and non-fire) that I used to feed the models.

• Metamodel
  • To construct the metamodel, I extracted the predictions from the topographic and imagery datasets, used them as features, and ran the two through a scikit-learn LogisticRegression, which achieved over 99% accuracy. This concluded the modeling portion of the project.

Findings and Conclusions

• The statsmodels Logit model reveals that higher elevations and west- and north-facing hillslopes correlate with wildfires, potentially due to dominant wind directions in the area.
• The project serves as a proof-of-concept, demonstrating the capabilities of a machine-learning model using remotely-sensed data for predicting wildfire risk. Future iterations can expand by incorporating weather and time series analysis and testing the model on different landscapes.
• Despite limitations, such as lack of high-resolution weather data, the project shows promise for developing a robust wildfire risk prediction model with broader data and automated GIS workflows.
• Publishing the model in a more accessible manner could provide communities with valuable insights into their wildfire risk.

Postscript: Commentary on why the accuracy is so high

• Accuracy is almost 100%, which is suspiciously high
• Since the model was trained on either 100% burned areas or 100% unburned areas, it only knows the clearly-delineated cases.
• Furthermore, when tested, the model was only given 100% burned or unburned areas
• The unburned area was a city center. There would never be a wildfire in a concrete jungle
• The opposite is similarly always true: in the mountains, away from all civilization, fires are extremely likely and are much harder to combat

Extra goodies

• Correlation matrix of the geographic data:

• Summary of models and results for TopoData:

• Summary of satellite imagery model:

• Summary of metamodel:

• Satellite image and historic wildfire boundaries (up to 2018):

• Satellite image and hillshade of study area:

Footnotes

1. [https://www.fire.ca.gov/incidents/2020/] ↩

2. [https://report.ipcc.ch/ar6syr/pdf/IPCC_AR6_SYR_SPM.pdf] ↩

3. [https://earthexplorer.usgs.gov] ↩

4. [https://data-nifc.opendata.arcgis.com/] ↩

5. [https://bigearth.net/] ↩