Utah Vehicle Accidents Prediction

Can we predict car accident risk on individual road segments in Utah? Let's do our best using the power of ArcGIS and Python!

For questions about these notebooks, or to obtain the data needed to run them, please contact Daniel Wilson

Approach

We are approaching this from a supervised machine learning perspective. Using a training set of hours where accidents occurred, augmented by a sample of times/roads where they didn't, we are training a gradient boosting model with a cross-entropy loss function (logistic regression) to estimate the accident risk. Note that I say risk - we are not estimating the probability directly, but the risk is proportional to the probability. This issue stems from class imbalance: There are about 400k accidents, but when looking at all the segments and times where they didn't occur, we're now talking about 200M+ training examples. That's what we call a severe class imbalance issue. We keep the training set slightly imbalanced, and use an informative sampling approach to get the model to learn the fine grained differences between what causes a car accident or not.

The Data

We have 7 years from car crash records for the state of Utah, but that's only the start of our training set for this supervised machine learning problem. We use many different features to try to differentiate between car accidents occuring or not. These include:

• Road Features
  • Curvature (Sinuosity)
  • Orientation (East/West vs North/South)
  • Speed Limit
  • Surface Type
  • Surface Width
  • Annual Average Daily Traffic (AADT)
  • Proximity to nearest intersection
  • Proximity to nearest signal
  • Proximity to nearest major road
  • Proximity to nearest billboard
  • Population density
  • Historical accident counts
• Weather Features
  • Temperature
  • Visibility
  • Precipitation/Snow Depth
  • Icy
  • Raining
  • Snowing
  • Hailing
  • Thunderstorming
• Temporal Features
  • Solar Azimuth/Elevation
  • Time of day
  • Day of the week
  • Month

There are more that we could add, and likely will as we refine this process.

We use ArcGIS Pro and Arcpy to perform some geoprocessing, followed by the ArcGIS API for Python, Pandas, and scikit-learn to prep the data. We also use some miscellaneous libraries along the way, such as Astral for solar geometry computation.

The weather data is pulled from the NOAA's API and aggregated into hourly bins per weather station. Currently, we chose to snap each road to the nearest weather station. We realize this is a poor approach, and will implement interpolation/regression to accurately asses weather conditions for each road. We would also like to bring in better weather sources, as well as real time weather forecasting as we build this into a more realtime application for situational awareness.

Any categorical variables are one-hot encoded using panda's pd.get_dummies and all continuous variables are converted to a z-score using scikit-learn's StandardScaler.

The Model

We use XGBoost library for the supervised machine learning model. We also experimented with Deep Neural Networks, but found gradient boosting to work better. The model is trained using a max_depth of 6, min_child_weight of 5.0, and a L2 penalty term (reg_lambda) of 1.0. We train until completion of best accuracy on a validation set with 25 rounds of early stopping.

The Notebooks

• prep_static_features.ipynb - ArcPy preparation of the road features
• get_weather.ipynb - pulls/transforms the NOAA weather data
• prepare_training_data.ipynb - builds a training set CSV file
• train_model.ipynb - trains the model and plots results.