Rapid Loss of the Sierra Nevada's Largest Trees Driven by Fire
Fabien H. Wagner, Dan J. Dixon, Christopher W. Woodall, Mayumi C. M. Hirye, Felipe Saad, Griffin Carter, Ricardo Dalagnol, Lorena Alves, et al.
cs.CV
Quick verdict
Wall-to-wall map of 6.5M Sierra Nevada large trees; fire killed ~10% in 2020-21. Compelling regional advance; validation transfer and uncertainty need scrutiny.
Summary
The authors combine a canopy height model derived from 2020 NAIP sub-meter aerial imagery with a U-Net-ID segmentation model trained entirely on 3.36 million synthetic tree crowns to delineate every tree with crown area of at least 100 m2 across the Sierra Nevada Floristic Province. Validation against 20,273 LiDAR-anchored giant sequoia crowns yields over 99.99% detection, median IoU 0.602 after centroid alignment, and a systematic -60.26 m2 area bias. They map 6,515,705 large trees across 78.7% of the province and relate density to elevation, precipitation, and temperature. For each crown they extract 2020-2025 Sentinel-2 NDVI/RGB trajectories, apply BFAST breakpoint detection, and classify crowns into eight health states; ground validation on 1,015 sequoias shows 96.7% agreement for the alive class but only 68.9% for the dead class. Overlapping the classifications with CAL FIRE fire perimeters and correcting with the observed error rates, they estimate fire killed about 660,000 trees (10.1%), concentrated in the 2020-2021 fire seasons, versus 1.6% mortality outside perimeters; burned sequoia groves show 17% mortality, close to a field-based 17.6% since 1984. Main limitations: error rates validated only on giant sequoias in three fire footprints are applied region-wide, fire attribution rests on spatial coincidence with coarse perimeters, the headline estimate lacks propagated uncertainty, and code and processed data are available only on request.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A U-Net-ID trained solely on 3.36 million synthetic crowns delineated 6,515,705 large trees (crown >=100 m2) wall-to-wall across the Sierra Nevada with >99.99% detection of reference sequoias and median IoU 0.602.
- Per-crown Sentinel-2 NDVI/RGB trajectories analyzed with BFAST classify individual-crown health, agreeing with field assessment for 96.7% of alive and 72% of strong-mortality crowns.
- Wildfire was the dominant driver of large-tree mortality in 2020-2025, killing an estimated ~660,000 trees (~10.1% of all mapped large trees), with mortality concentrated in the 2020-2021 fire seasons (87.9% of burned-area observations).
- Within burned giant sequoia groves, ~17.0% of large trees died, consistent with the field-based estimate of 17.6% killed by fire since 1984.
- Large-tree occurrence is concentrated at 600-1,700 m elevation, above ~800 mm annual precipitation, and 16-22 C mean maximum temperature.
Strengths
Regional scale bridging the resolution gap
Wall-to-wall delineation of 6,515,705 individual crowns (>=100 m2) across 65,582 km2 from 2020 NAIP-derived CHMs, then per-crown Sentinel-2 trajectories - directly addressing the spatial-temporal trade-off between static VHR mapping and medium-resolution time series set out in the Introduction.
Fully synthetic training pipeline
3,356,877 simulated crowns in 59,134 CHM patches remove manual annotation dependence; sim-to-real transfer yields median IoU 0.602, above the 0.53 reported by Cheng et al. [28] for dead-tree crowns on related California imagery.
Error-aware mortality accounting
Independent plausibility cross-check
Transparent geolocation handling
Weaknesses
Sequoia-only validation extrapolated region-wide
Fire attribution by perimeter overlay only
No uncertainty quantification
Validation reference is itself an estimate with temporal mismatch
Reproducibility limited to text
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How many large trees are in the Sierra Nevada, and how many were lost to wildfire?
This study mapped 6,515,705 individual trees with crown areas of at least 100 square meters across the Sierra Nevada Floristic Province using 2020 sub-meter aerial imagery and deep learning. It estimates that about 660,000 of them - roughly 10% - were killed by wildfire between 2020 and 2025, with most of that mortality concentrated in the extreme 2020 and 2021 fire seasons.
How did researchers detect and monitor individual trees from imagery?
The team built a canopy height model from sub-meter NAIP aerial imagery, then ran a U-Net deep learning model trained entirely on 3.3 million simulated (synthetic) tree crowns to segment each tree without manual labels. Each detected crown's 2020-2025 Sentinel-2 NDVI and color time series were then analyzed with the BFAST algorithm to detect disturbance, death, or recovery. Detection of known giant sequoias exceeded 99.99%, though median spatial overlap with reference crowns was moderate (IoU 0.602).
Are giant sequoias dying from California wildfires?
Substantially, yes. The study estimates about 17% of large trees within burned sequoia groves died in the 2020-2021 fires, consistent with field research finding that 17.6% of large giant sequoias have been killed by wildfire since 1984, with roughly 80% of those deaths in 2020-2021. Survival was heterogeneous: many large crowns classified as alive persisted inside fire perimeters, reflecting unburned patches and fire resistance.
How accurate is the satellite-based tree health classification?
Against 1,015 field-checked giant sequoias, the alive class was 96.7% accurate and the strongest mortality class 72% accurate; the explicit dead class agreed with field status in 68.9% of cases, and 38% of crowns showing strong post-fire NDVI recovery were actually dead, because greening can come from neighboring trees or understory. The authors used these observed error rates to adjust their mortality estimate downward before reporting the ~10% regional loss.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Map figures, segmentation examples, and time-series panels could not be visually assessed: the text extraction garbles figure content (e.g., the Fig. 6b legend repeats dozens of times; Fig. 12 axis labels are scattered).
- Minor internal inconsistencies suggest light copyediting: validation n given as 20,273 throughout but 20,723 in the Discussion; elevation floor 150 m (Introduction) versus 30 m (Methods 4.1); 80% versus >=85% NDVI-recovery threshold between text and Fig. 12 labels.
- Review is based on the arXiv v1 preprint (15 Sep 2026, cs.CV); supplementary material referenced in the text was not included in the extraction.
- Field ground truth covers only the Castle, Windy, and KNP Complex fire footprints (1,015 of 3,036 monitored sequoias intersected), so per-class error rates rest on small subsamples.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.