Georisk Continuous (Generalized), Santa Rosa, CA

The dataset shows the risk of a water sample exceeding the Maximum Contaminant Level (MCL) after a wildfire.

Grid cell values provide absolute probabilities of drinking water contamination above the State of California MCL for Benzene (1 µg/L) for water samples from the water distribution system after a collocated wildfire event.

Bayesian regularized neural network ensembles were trained using high-resolution data layers comprising topography, soil properties, landcover, vegetation, meteorological parameters, fuel load, and infrastructure data. In combination with post-fire water samples, the input data was used to map the risks of MCL exceedance to the values observed in the cities of Santa Rosa, CA and Paradise, CA.

Many contributing factors and processes that can cause the post-fire contamination of drinking water in water distribution systems (WDS) are partially unknown or corresponding data unavailable. Processes such as the water distribution system-wide state of pressure, flow, and temperature in a complex pipe network across a town are unknown, except for certain main valves and control points.

Furthermore, parameters change during wildfires when firefighting efforts or damaged pipes and associated pressure drops change flow rates and directions at one or many points of the distribution system.
Furthermore, current wildfire models do not allow for modeling burn probability or fire behavior in built-up areas due to a current lack of fuel models for such structures. Sections of built-up areas containing numbers of structures that can be close to burnable vegetation are currently classified as non-burnable in fuel layers of fire models. Hence, using a deterministic process model for spatial predictions of post-fire contamination risk with available sampling data and knowledge of processes, is currently unfeasible. For the spatial analyses here, we use a machine learning approach with pattern recognition networks that have SoftMax classification output layers to spatially predict conditional probabilities of drinking water contamination in WUI areas after fire affected the structures and the surrounding areas. We use analytical results of post-fire water samples, topographic factors, landcover data, information about infrastructure, and physical soil properties in combination with Bayesian regularized neural networks building ensemble models that predict conditional probabilities for benzene levels in WDS exceeding the maximum contaminant level (MCL) for benzene. Benzene is considered a carcinogen and poses a severe health threat to humans if consumed in high concentrations. While other contaminants were found in WDS water samples after wildfires, benzene was chosen as a representative Volatile organic compound because of its abundance in post-fire water samples in Santa Rosa and Paradise, California.

Using the water samples that were collected in any study area after the wildfire, the parameters of the neural networks are iteratively optimized to map the input data on the target data (i.e., the contamination status of post-fire water samples at each point).

Once the model is optimized to reproduce the training data and generalize well enough to also model new data (not included in the training process) with sufficient accuracy, the model is applied to the entire model domain with a 30 m x30 m resolution as illustrated in figure 1. Several models can be averaged to build an ensemble result at each grid cell point which in practice often increases the generalization capabilities of the models and hence, their accuracy. The results for the risk of water contamination shown as part of the EEMS model give the conditional probability that post fire water samples exceed the California MCL for benzene in drinking water (1 µg/L) after a potential fire.