Sentinel imagery detects the presence of live trees following large wildfires in California

Identifying live tree presence following wildfire is important for burn damage assessments and decision making, as these trees serve as seed sources for recovery. Satellite-based remote sensing offers an efficient means to assess burn severity with products representing vegetation greenness and char/ash presence and their change from pre- to post-fire imagery. While effective at assessing burn severity (e.g. ecosystem change), there remain limitations in identifying fire refugia (surviving trees), due to the difficulty of teasing apart different green vegetation types (e.g. trees, shrubs, grasses). In this paper, we use 10 m Sentinel-2 satellite data to predict live tree presence across three sites impacted by the 2021 California fire season. We used vegetation indices (VIs) from post-fire imagery (normalized difference vegetation index [NDVI], normalized burn ratio [NBR], normalized difference water index [NDWI], visible atmospherically resistant index [VARI], and burn area index [BAI]), differential VIs from pre- and post-fire imagery (dNDVI, dNBR, RdNBR, dNDWI, dVARI), and direct reflectance bands (all bands model; visible, near-infrared, and shortwave infrared; B1–B12) to predict live tree presence via random forest modeling. To calibrate and validate the random forest models, we photointerpreted ∼2300 pixels per fire region using 2022 National Agriculture Imagery Program imagery. We performed additional field-based validation using tree presence/absence data two years post-fire (n = 296 observations across two sites). At the site level, the all bands model outperformed the vegetation index-based models (80%–85% vs 65%–79% accuracy). Errors were mainly false positives attributed to pixels with green understory vegetation but no live trees. In cross-site inference, which involved pooling two sites for model calibration to test on the third site, the all bands model retained good performance (76%–81% accuracy). Evaluation against field survey data demonstrated a larger range of performance (50%–87% accuracy) that highlights limitations based on tree isolation and crown percent greenness. Relative to differential-based VIs, our results highlight potential advantages of using post-fire Sentinel-2 imagery and random forest modeling for identifying live tree presence and scaling to full fire extents.