cs.CVSep 15, 2026
SaveRapid Loss of the Sierra Nevada's Largest Trees Driven by Fire
Organizations: CTrees, Pasadena, CA, 91105, USA. · Department of Land, Air and Water Resources, University of California, Davis, Davis, CA, 95616, USA. · Institute of Environment and Sustainability, University of California, Los Angeles, Los Angeles, CA, USA. · Quap´a Lab, Faculty of Architecture and Urbanism, University of S˜ao Paulo, 05508080, S˜ao Paulo, SP, Brazil. ·
Abstract
Large trees disproportionately contribute to biomass storage, habitat structure, and ecosystem functioning. However, their distribution and health dynamics remain poorly quantified at a regional scale. Here, a deep learning model (U-Net-ID) and canopy height models derived from sub-meter aerial imagery from 2020 were used to delineate all individual trees with crown area 100 m across the Sierra Nevada Floristic Province. The model was trained using more than 3.3 million synthetic tree crowns and achieved a median Intersection over Union (IoU) of 0.602 when validated against an independent dataset of 20,273 crowns. A total of 6,515,705 large trees were mapped, occurring across approximately 78.7% of the Sierra Nevada Floristic Province. The spatial distribution of large trees showed associations with elevation, temperature, and precipitation. Using Sentinel-2 time series from 2020 to 2025, tree health dynamics were characterized by extracting spectral trajectories for each crown and applying BFAST breakpoint detection algorithm combined with a disturbance classification framework to identify mortality, disturbance, and recovery trajectories of individual trees. Wildfires, estimated from CAL FIRE fire perimeters, were identified as the dominant driver of large-tree mortality, killing 10% of all large trees in the Sierra Nevada, with mortality strongly concentrated during the extreme 2020-2021 fire seasons.
Figures & tables
Figure 1: Comparison of predicted versus observed crown area (m 2 ) for the 20,273 Sequoia trees with observed polygons delineated manually from CHM derived from LiDAR 2015/2016 data based on field GPS coordinates (1 polygon had no intersection). The 1:1 line is depicted in dashed gray. A tendency curve based on LOESS smoothing is shown to illustrate the relationship between observed and predicted crown area.
Figure 2: Example of validation images for giant sequoia trees with manual delineation in blue and the U-Net-ID-based delineation in red.
Figure 3: Distance between the predicted and observed tree polygon centroids for the 20,272 Sequoia trees with a predicted polygon (a), and distribution of the Intersection over Union (IoU) between predicted and observed Sequoia polygons after centroid alignment (n=20,272) (b). Centroids were aligned to account for small geolocation errors between intersected predicted and observed polygons.
Figure 4: Distribution of the 6,515,705 tree crowns 100 m 2 delineated with the U-Net-ID model in the Sierra Nevada.
Figure 5: Proportion of all Sierra Nevada pixels or pixels containing large trees across classes of environmental and climatic variables: (a) elevation (m), (b) slope ( ∘ ), (c) slope orientation (North, East, South, West), (d) mean annual precipitation (mm yr -1 ), (e) annual mean minimum temperature ( ∘ C), and (f) annual mean maximum temperature ( ∘ C). For elevation and precipitation, class values on the x-axis represent class means to ease visualization. Values of the ratio of observed (large trees) over expected (available area) equal to 1 indicate a uniform representation, values above 1 indicate over-representation of large trees, and values below 1 indicate under-representation.
| Ground based observations | ||
|---|---|---|
| Alive | Dead | |
| Satellite based observation | ||
| alive (1) | 202 (96.7%) | 7 (3.3%) |
| alive_with_bkpt (2) | 142 (95.3%) | 7 (4.7%) |
| bkpt_ndvi_sup80 (3) | 57 (62.0%) | 35 (38.0%) |
| bkpt_ndvi_inf80 (4) | 142 (28.0%) | 365 (72.0%) |
Table 1: Ground-based observations of giant sequoia mortality compared to our remotely sensed classification of tree crown ( 100 m 2 ) status classes for the 1,015 intersecting trees between both datasets.
Figure 6: Alive, disturbed and dead status classes of tree crowns 100 m 2 in the Sierra Nevada with similar spectral characteristics as the sequoias (a). To ease visualization, pixel values represent the most represented class at 500 m spatial resolution. (b) Fire occurrence during the 2020-2025 period in the Sierra Nevada at 500 m spatial resolution for the pixel that contains at least one tree crowns 100 m 2 . The names of the eight largest fires, that all occurred in 2020 and 2021, are displayed.
| Class | Class ID | Burned (N, %row, %col) | Unburned (N, %row, %col) | Total (N) |
|---|---|---|---|---|
| alive | 1 | 271755 (7.23%, 17.63%) | 3486934 (92.77%, 70.09%) | 3758689 |
| alive_with_bkp | 2 | 152017 (18.71%, 9.86%) | 660432 (81.29%, 13.28%) | 812449 |
| bkp_sup80 | 3 | 152543 (59.61%, 9.9%) | 103346 (40.39%, 2.08%) | 255889 |
| bkp_inf80 | 4 | 692992 (92.38%, 44.97%) | 57156 (7.62%, 1.15%) | 750148 |
| dead | 6 | 146775 (79.61%, 9.52%) | 37591 (20.39%, 0.76%) | 184366 |
| permanent_alteration | 5 | 20576 (37%, 1.34%) | 35029 (63%, 0.7%) | 55605 |
Table 2: Large tree crown health status estimated from remote sensing and their distribution within CAL FIRE fire perimeters.
| Class | Class ID | Burned (N, %row, %col) | Unburned (N, %row, %col) | Total (N) |
|---|---|---|---|---|
| alive | 1.00 | 18346 (50.92%, 41.57%) | 17680 (49.08%, 76.47%) | 36026 |
| alive_with_bkp | 2.00 | 7096 (65.67%, 16.08%) | 3710 (34.33%, 16.05%) | 10806 |
| bkp_sup80 | 3.00 | 2682 (87.05%, 6.08%) | 399 (12.95%, 1.73%) | 3081 |
| bkp_inf80 | 4.00 | 12223 (98.38%, 27.69%) | 201 (1.62%, 0.87%) | 12424 |
| dead | 6.00 | 2330 (96.24%, 5.28%) | 91 (3.76%, 0.39%) | 2421 |
| permanent_alteration | 5.00 | 831 (86.47%, 1.88%) | 130 (13.53%, 0.56%) | 961 |
Table 3: Large tree crown health status estimated from remote sensing within the sequoia groves and their distribution in CAL FIRE fire polygons.
Figure 7: Spatial pattern of large tree disturbance in the North Complex Fire. Only the main live/disturbed health classes are shown to ease visualisation.
| Class | Class ID | Burned (N, %col) |
|---|---|---|
| alive | 1 | 51429 (13.09%) |
| alive_with_bkp | 2 | 54832 (13.95%) |
| bkp_sup80 | 3 | 69085 (17.58%) |
| bkp_inf80 | 4 | 188490 (47.97%) |
| dead | 6 | 21758 (5.54%) |
| permanent_alteration | 5 | 4211 (1.07%) |
Table 4: Large tree crown health status estimated from remote sensing within the largest extent of the North Complex Fire in December 2020.
Figure 8: Mosaic of NAIP aerial imagery acquired over California (USA) during the 2020 campaign, with the Sierra Nevada Floristic Province outlined in dashed yellow ( 65,582 km 2 ). The map also shows the extent of the 22 Sentinel-2 MSI tiles used for time-series analysis (grey) and the extent of the 20,273 Sequoia polygons used for individual tree crown model validation (magenta).
| Scenario (short name) | range | con. | t.d. | l.c.d. | m.d. | s.c. | s.d. | sh.r. | sh.s. | g.d. | g.c. | g.r. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| dense_redwood | 20–30 | .05 | .05 | .05 | .05 | .05 | .05 | .05 | .05 | .10 | .10 | .40 |
| redwood_conif_mix | 25–35 | .20 | .05 | .05 | .05 | .10 | .05 | .05 | .05 | .05 | .15 | .20 |
| old_growth_redwood | 20–30 | .10 | .10 | .10 | .10 | .05 | .05 | .05 | .05 | .10 | .10 | .20 |
| redwood_giant_mix | 25–35 | .10 | .10 | .10 | .10 | .05 | .05 | .05 | .05 | .15 | .15 | .15 |
| pure_redwood | 15–25 | .05 | .05 | .05 | .05 | .05 | .05 | .05 | .05 | .05 | .05 | .50 |
| redwood_tall | 25–35 | .10 | .20 | .10 | .10 | .05 | .05 | .05 | .05 | .10 | .05 | .15 |
Table 5: Example of giant sequoia forest scenarios. For each scenario, the number of trees per simulation is drawn uniformly from the integer range in the second column. The remaining columns are sampling weights for functional types (row sums to 1).
| Type (code name) | Height (scaled) | Height (m) | Crown radius (px) | Crown radius (m) | |
|---|---|---|---|---|---|
| coniferous | 50–100 | 20.0–40.0 | 4–6 | 2.4–3.6 | 12 |
| tall_deciduous | 60–100 | 24.0–40.0 | 3–5 | 1.8–3.0 | 15 |
| large_crown_deciduous | 35–80 | 14.0–32.0 | 6–8 | 3.6–4.8 | 6 |
| medium_deciduous | 40–70 | 16.0–28.0 | 4–6 | 2.4–3.6 | 9 |
| sapling_coniferous | 10–25 | 4.0–10.0 | 1–2 | 0.6–1.2 | 10 |
| sapling_deciduous | 8–20 | 3.2–8.0 | 1–3 | 0.6–1.8 | 6 |
Table 6: Tree functional types and their height and crown parameters used in the CHM simulation. Height values in the simulator are stored in scaled units; physical height (m) is obtained by dividing by . Crown radius is given in pixels in the simulator; crown radius (m), where one pixel corresponds to m on the ground. The column corresponds to the height_radius_ratio field in the same list.
Figure 9: Example of simulated individual tree polygons and the associated simulated CHM, simulated blurred CHM and masks in raster format.
Figure 10: U-Net-Id model architecture used for individual tree segmentation from Canopy height model, adapted from Wagner et al. Wagner et al. (2020) . The three U-net units used in our model have the same layers as in Wagner et al. Wagner et al. (2020) only that the input and output of the U-net units here are 128 128. The outputs of the U-net units are represented with three different colors: magenta (U-net 1 ), blue (U-net 2 ), and green (U-net 3 ). The number of channels is indicated above the cuboids, and the vertical numbers indicate the row and column size in pixels. The operations (skip connections and convolutions) performed in each layer and their sizes are indicated by the colored arrows.
Figure 11: Sample of points on the surface defining the RGB color domain from sequoia field crowns extracted from Sentinel-2 imagery (a), and a 3D mesh representing the color-domain surface of sequoia crowns used to filter the Sentinel-2 time series, i.e., all colors outside the mesh are discarded (b).
Figure 12: Examples of the seven time series trajectories identified in the dataset and used to classify the large crowns into alive or disturbed classes.
Figure 13: Environmental and climatic variables used in the study to analyse large trees density distribution (a) elevation (m), (b) slope ( o ), (c) Slope orientation in classes (North, East, South and West), (d) mean annual precipitation (mm.yr -1 ), (e) annual mean of minimum temperatures ( o C), and (f) annual mean of maximal temperatures ( o C).