Scalable sub-meter mapping of woody vegetation and structures across California’s heterogeneous landscape using deep learning

The encroachment of human settlements into wildland areas has expanded the Wildland-Urban Interface (WUI), disrupting ecosystems and increasing wildfire risks. A systematic mapping of woody vegetation and building structures across the wildland-urban continuum is critical for fire risk and ecosystem service assessment, fuel management, conservation, and land use planning. The highly heterogeneous and dynamic nature of this landscape requires a fine-scale monitoring approach. Although machine learning and computer vision have advanced land cover classification using very high resolution imagery (VHR), most studies either focus on identifying only a particular object or a limited study area. We here developed a unified semantic segmentation deep learning model (U-Net) to map trees, shrubs, and building footprints at 0.6 m resolution using the open-source VHR multispectral imagery from the National Aerial Imaging Program (NAIP) for the state of California. A semi-automatic labeling process was adapted to generate a large set of labels for model training and testing with available aerial LiDAR surveys. The validation showed robust performance of the trained U-Net model in predicting canopy-scale trees and shrubs as well as buildings, achieving an overall accuracy of 87.1% and F1 scores of 83.1% for trees and 78.9% for shrubs across different years. Further evaluation demonstrated that this approach captured the fine-scale spatial arrangement of woody vegetation and buildings, and the temporal vegetation dynamics from selective logging, regrowth, and tree mortality following wildfire. The model’s scalability was also shown for county and statewide mapping. Given the availability of open-access NAIP imagery every 2–3 years over the continental US, our scalable approach provides a sub-meter monitoring tool and data to improve ecological and building assessment and fire simulation. These capabilities support adaptive land management to mitigate WUI fire risk and ultimately promote fire-safe communities and resilient ecosystems along the WUI-wildland gradient. • Unified U-Net model maps woody fuels and buildings across the urban-wildland continuum. • Semi-automatic workflow expedites label collection from aerial LiDAR and NAIP images. • Robust performance in tracking fine-scale fuel dynamics across California. • Demonstrates potential for tracking post-fire housing loss and rebuilding.