henhao
Sign in
EcologyForestry An International Journal of Forest Research
Peer-reviewed journal article

Large-scale remote sensing reveals that tree mortality in Germany appears to be greater than previously expected

Felix SchieferiD, Sebastian SchmidtleiniD, Henrik HartmanniD, Florian SchnabeliD, Teja KattenborniD

Abstract

Global warming poses a major threat to forests and events of increased tree mortality are observed globally. Studying tree mortality often relies on local-level observations of dieback while large-scale analyses are lacking. Satellite remote sensing provides the spatial coverage and sufficiently high temporal and spatial resolution needed to investigate tree mortality at landscape-scale. However, adequate reference data for training satellite-based models are scarce. In this study, we employed the first maps of standing deadwood in Germany for the years 2018-2022 with 10 m spatial resolution that were created by using tree mortality observations spotted in hundreds of drone images as the reference. We use these maps to study spatial and temporal patterns of tree mortality in Germany and analyse their biotic and abiotic environmental drivers using random forest regression. In 2019, the second consecutive hotter drought year in a row, standing deadwood increased steeply to 334 {+/-} 189 kilohectar (kha) which corresponds to 2.5 {+/-} 1.4% of the total forested area in Germany. Picea abies, Pinus sylvestris, and Fagus sylvatica showed highest shares of standing deadwood. During 2018-2021 978 {+/-} 529 kha (7.9 {+/-} 4.4%) of standing dead trees accumulated. The higher mortality estimates that we report compared to other surveys (such as the ground-based forest condition survey) can be partially attributed to the fact that remote sensing captures mortality from a birds eye perspective and that the high spatial detail (10 m) in this study also captures scattered occurrences of tree mortality. Atmospheric drought (i.e., climatic water balance and vapor pressure deficit) and temperature extremes (i.e., number of hot days and frosts after vegetation onset) were the most important predictors of tree mortality. We found increased tree mortality for smaller and younger stands and also on less productive sites. Monospecific stands were generally not more affected by mortality, but only when interactions with damaging insects (e.g., bark beetles) occurred. Because excess tree mortality rates threaten many forests across the globe, similar analyses of tree mortality are warranted and technically feasible at the global scale. We encourage the international scientific community to share and compile local data on deadwood occurrences (see example: www.deadtrees.earth) as such a collaborative effort is required to help understand mortality events on a global scale.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
22
FWCI
3.45
Field percentile
top 6%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (11 Nov 2024) · CC BY 4.0

Germany-wide 10 m remote sensing maps put 2018–2021 standing deadwood at 978 ± 529 kha, almost twice earlier satellite estimates; late frost, water deficit and heat were top predictors.

Landscape-scale studies of tree mortality have been held back by small-sample ground surveys and by satellite products that are coarse (MODIS 250 m, Landsat 30 m) or use vegetation indices as a stand-in for dieback. This work instead relates Germany-wide 10 m maps of directly estimated standing deadwood, trained on drone-image segmentations, to tree species and environmental drivers.

Key results

  1. ShownNew standing deadwood rose sharply in 2019, the second drought year in a row. By 2021 it had accumulated to 7.9 ± 4.4% of German forest area, and the conifers Picea and Pinus were hit hardest.

    2018: 179 ± 126 kha (1.4 ± 1.0%); 2019: 334 ± 189 kha (2.5 ± 1.4%); 2020: 307 ± 151 kha (2.6 ± 1.3%); 2018–2021 total: 978 ± 529 kha. Accumulated share by species: Picea 21.5 ± 7.1%, Pinus 20 ± 12%, Fagus 8.2 ± 3.8%, Quercus 4.5 ± 1.7%. Most affected region by share: Harz (D37, 30.2%). The ± values are the SD across first-occurrence thresholds of 30–70%. · Results; Tree mortality in Germany

  2. SuggestedEstimates are overall higher than the ground-based forest condition survey (clearest in 2018–2019 and for species totals; in 2021 the survey reported more, 2.35% vs 1.3 ± 0.6%) and almost twice two satellite loss products. The authors credit 10 m detail and overhead view; no ground truth tests which is right.

    Survey vs this study: 2018, 0.51% vs 1.4 ± 1.0%; accumulated Pinus 3.8% vs 20 ± 12%, Picea 10.4% vs 21.5 ± 7.1%, Fagus 2.4% vs 8.2 ± 3.8%. Area for 2018–2021: 978 ± 529 kha vs 501 kha (Thonfeld et al. 2022) and 543 kha (Global Forest Watch). Only 26.6 ± 14.5% of deadwood pixels were also flagged as forest loss by Global Forest Watch. · Pattern of tree mortality in Germany; Conclusion

  3. ShownIn the per-species random forests, the strongest predictors were atmospheric conditions (late frost, climatic water balance, hot days, vapour pressure deficit) plus canopy height. Soil, productivity, stand age and species richness ranked lower.

    Median permutation importance (RMSE) across seven species: late frost 15.3%, canopy height 15.2%, climatic water balance 14.3%, hot days 13.7%, vapor pressure deficit 12.1%; soil drought intensity 9.8%, sand 9.4%, stand age 8.3% (rank 17), species richness 7.3% (rank 19). Model R² ranged from 0.36 (Larix) to 0.68 (Pinus). Moran's I of the residuals was close to 0 (± 0.04). · Results; Identifying predictors of tree mortality using…

  4. ShownAccumulated local effects link lower canopy height, lower productivity and biomass, and slightly younger stands to more deadwood, and drought thresholds differ by species. The authors read the canopy-height effect as a sign that open stands and forest edges are vulnerable.

    Smaller canopy height went with more deadwood in all seven species (largest effect for Larix, smallest for Abies). Deadwood rose above a vapor pressure deficit of 2.1 kPa for Pinus vs 1.5-1.7 kPa for Quercus, Larix, Pseudotsuga and Picea; Fagus and Abies showed the reverse. It also rose above 30% sand content for Fagus and Abies vs 75% for the other species. · Results; Canopy height; Site productivity and biomass

  5. SuggestedTree species richness mattered little overall, and its effect depended on species. More diverse stands had more deadwood for Fagus and Quercus but less for Picea. The authors attribute the Picea pattern to bark beetles in monocultures.

    Richness ranked 19th (7.3%). ALE: no clear effect for Larix, Pseudotsuga, Pinus and Abies; positive association for Quercus and Fagus; negative for Picea. The species map records only the dominant species per pixel. · Results; Tree species and species richness

Methods, data and reproducibility

Earlier Germany-wide 10 m deadwood cover maps came from convolutional networks on drone images, upscaled by LSTM networks on Sentinel-1/2 (RMSE 22%). Each pixel was reduced to its first year above a 30–70% cover threshold and overlaid with a 10 m species map and 73 natural regions. Per-species random forests (100 repeats, permutation importance, ALE) related deadwood to 21 environmental predictors.

  • Germany-wide maps of standing deadwood fractional cover, 10 m, 2018–2021 (Schiefer et al. 2023; drone-derived training data, Sentinel-1/2 predictors) · forest trees
  • Dominant tree species map of Germany, 10 m (Blickensdörfer et al. 2024), based on national forest inventory data and remote sensing · trees (e.g., Picea abies, Pinus sylvestris, Fagus sylvatica) · n = 11 tree species
  • Pixel samples for per-species random forests (half with, half without standing deadwood; 4,000 per year of first occurrence) · 7 tree species: Fagus, Quercus, Picea, Pinus, Abies, Pseudotsuga, Larix · n = 32,000 pixels per model, 100 repetitions
  • Environmental predictors: climate/weather, soil, productivity, stand structure, topography (Table 1) · n = 21 variables
  • Major natural regions of Germany (Meynen and Schmithüsen 1953, adapted by BfN) · n = 73 regions
  • German national forest condition survey (BMEL, 2023), used for comparison · trees · n = 9688 trees from 402 sample plots
  • Forest-loss products used for comparison: Thonfeld et al. (2022) and Global Forest Watch
  • Data underlying the article (KITopen repository) · 10.5445/IR/1000158765
Methods & tools
R, ranger, iml, random forest, accumulated local effects (ALE), permutation importance, Moran's I, Sentinel-1, Sentinel-2, convolutional neural networks (upstream map), LSTM (upstream map), Global Forest Watch
Code
https://github.com/FelixSchiefer/TreeMortalityPattern
Data & accessions
https://dx.doi.org/10.5445/IR/1000158765https://deadtrees.earth/
Software & environment
Analyses were run in R with the ranger (random forest) and iml (ALE) packages; package versions are not stated.
Reproducibility gaps
R package versions and compute are not reported. Predictor definitions (Table 1), the appendix on variable selection and supplementary tables S2–S3 are referenced but not included in the text.
Limitations (6)
  • The underlying deadwood map sometimes overpredicts near forest edges, roads and paths. Forest floor showing through sparse Pinus canopies (after defoliation or insect damage) may be read as deadwood, which shows up as a large threshold-sensitivity SD for Pinus. · stated by the authors
  • Abies and Pseudotsuga have the lowest accuracies in the species map and are often confused with Picea, so their deadwood is likely overestimated. · stated by the authors
  • Direct comparison with the forest condition survey is hampered: the survey counts individual trees, while the maps are area-based and also include partly dead crowns and branches. · stated by the authors
  • Purely observational: mechanisms and confounding of the richness effect cannot be tested, and the species map gives only each pixel's dominant species. The authors note that computing and model limits often reduce environmental variables to single time points (here current and prior year) instead of time series. · stated by the authors
  • The claim of higher-than-expected mortality rests on comparisons with other products, and no new field validation is presented. The input map has a per-pixel RMSE of 22%, and the national total carries ± 529 kha uncertainty on 978 kha. · noted in this summary
  • None of the 21 predictors measures insects, so the bark-beetle explanation for the Picea monoculture effect, and the abstract's statement about insect interactions, are interpretation rather than tested results. · noted in this summary

Who should read:Read this if you study drought-driven forest dieback, upscale drone-derived labels to satellite maps, or want species-resolved drivers of Central European tree mortality with open code and data.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGanz '25Changenet '20Mosig '26Neumair '22George '21Battison '24
70% semantic matchEcology
Keenan J. Ganz, Liz van Wagtendonk, Pratima Khatri‐Chhetri+1 more·bioRxiv·Nov 19, 2025· 1 citations