Machine Learning with Applications (Mar 2023)
Application of deep convolutional networks for improved risk assessments of post-wildfire drinking water contamination
Abstract
Climate change continues to increase the frequency of wildfires in the western United States, driven by land use change and prolonged and intensified droughts. Simultaneously, an ongoing extension of communities into wildland areas has been observed over the last decades, increasing the wildfire threat to populations in the wildland–urban interface. In recent years, the problem of increased levels of volatile organic compounds in drinking water distribution systems after wildfires has gained attention as it poses a significant health threat to communities in those wildfire-prone areas. No adequate deterministic process models to predict the risk of high levels of volatile organic compounds in drinking water distribution systems after wildfires are available at this point, leaving data-driven machine learning approaches as solutions for addressing this problem. Here we build on a preceding study and enhance the assessment of maximum contamination exceedance probability predictions by applying deep convolutional neural networks trained with water samples from communities affected by wildfires in California and Oregon. We used satellite surface reflectance imagery in combination with gridded information of topography, fuel load, meteorology, and infrastructure as model inputs. The results show that the ensemble of deep convolutional neural networks increases the accuracy of predictions of post-wildfire water contamination by 4.1% compared to preceding models based on Bayesian regularized shallow feedforward networks, reaching an overall accuracy of 92% for the test dataset. The method provides a practical approach for emergency planning and pre-fire allocation of resources to support the provision of safe drinking water for wildfire-prone communities.
