Background: Wildland–urban interface (WUI) maps identify areas with wildfire risk, but they are often outdated owing to the lack of building data. Convolutional neural networks (CNNs) can extract building locations from remote sensing data, but their accuracy in WUI areas is unknown. Additionally, CNNs are computationally intensive and technically complex, making them challenging for end-users, such as those who use or create WUI maps, to apply.

Aims: We identified buildings pre- and post-wildfire and estimated building destruction for three California wildfires: Camp, Tubbs and Woolsey.

Methods: We evaluated a CNN-based building dataset and a CNN model from a separate commercial vendor to detect buildings from high-resolution imagery. This dataset and model represent to end-users the state of the art of what is readily available for potential WUI mapping.

Key results: We found moderate accuracies for the building dataset and the CNN model and a severe underestimation of buildings and their destruction rates where trees occluded buildings. The CNN model performed best post-fire with accuracies ≥73%.

Conclusions: Existing CNNs may be used with moderate accuracy for identifying individual buildings post-fire and mapping the extent of the WUI. The implications are, however, that CNNs are too inaccurate for post-fire damage assessments or building counts in the WUI.