Using remote sensing to support forest policies in Bavaria, Germany

Song, Qian; Ni, Xiaoshan

doi:10.5194/egusphere-2026-2782

Preprints

https://doi.org/10.5194/egusphere-2026-2782

Preprints

03 Jun 2026

| 03 Jun 2026

Status: this preprint is open for discussion and under review for Earth Observation (EO).

Using remote sensing to support forest policies in Bavaria, Germany

Qian Song and Xiaoshan Ni

Abstract. In order to mitigate climate change, the German Federal Government set goals to reduce greenhouse gas emissions and reach climate neutrality by 2045. To implement it, the federal and local governments have made a series of policies to improve the forest conditions in Bavaria. In this paper, we generated annual high-resolution dominant leaf type (DLT) and above-ground biomass density maps over Bavaria to support policy-making. Specifically, two U-Net-based models were trained to predict the DLT and biomass density separately from multispectral Sentinel-2 data based on deep learning. The model achieved 92.5 % DLT segmentation accuracy and an R² of 0.62 biomass estimation accuracy on the test set. Then, the trained model is used to derive annual DLT and biomass density maps from 2015 to 2025, where a post-processing step was proposed to exclude noisy fluctuating predictions. The results show a clear increase in tree area and broadleaved area, but this has slowed down since 2020. Besides, biomass loss due to tree degradation is higher than that due to deforestation, as suggested by the results. Subsequently, the time-series maps are used to identify hotspots in Bavaria, which is of interest to policymakers. We analyzed the tree cover and biomass loss for different administrative regions, and found that for most administrative areas, the increase of broadleaf tree areas is noticeably larger than the loss of that, except for Upper Franconia. Besides, continuous increases in both forest area and biomass amount in mountainous regions were observed. A landscape metrics-based analysis suggests that forest cover across the entire state has become increasingly fragmented. The results provide good insights into the tree status in Bavaria and suggest a new focus for forest management policies.

Received: 14 May 2026 – Discussion started: 03 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Qian Song and Xiaoshan Ni

Status: open (until 23 Sep 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-2782', Anonymous Referee #1, 07 Jul 2026 reply
For final publication, the manuscript should be reconsidered after major revisions. I would be willing to review the revised version.
On originality, I rate the paper Fair. A U-Net trained on Sentinel-2 for dominant leaf type and above-ground biomass is a well-established approach, and the authors themselves cite prior work along these lines (Waser et al., 2021; Song et al., 2024). The genuinely new part is the policy framing in Table 1 and the annual 2015 to 2025 time series for Bavaria, rather than the method. Relying on optical multispectral data alone for biomass, without any 3-D information (GEDI, LiDAR) or SAR (Sentinel-1), is also a fairly conservative choice given what is now available.
On scientific quality (rigour) I rate it Poor. Two issues are each serious on their own. First, the headline changes are smaller than the model's own uncertainty. Second, the resolution of the biomass reference and the saturation of the optical signal are not addressed. Taken together, they undercut most of the quantitative conclusions as they currently stand.
On significance, I rate it Fair. Linking the results to forest policy is worthwhile, but the high prediction uncertainty means the specific statements about how much biomass was gained or lost could mislead the very stakeholders the paper is written for.
On presentation, I rate it Poor. There are a lot of typos, at least one unresolved LaTeX citation, inconsistent handling of numbers and units, and several figures that are hard to read (overlapping legends, and different-unit metrics forced onto one rescaled axis).
General comments
My main concern is that the reported changes are smaller than the model's uncertainty. The central biomass claim is that the forest-average value "clearly increased" from 171.05 to 171.76 Mg/ha (l.277 to 278), which is about 0.4 %. That comes from a model with a test rRMSE of 0.45, i.e. roughly 45 % relative error, and the series is not even monotonic (Fig. 9 goes 171.05, up to 176.51 in 2021, down to 170.36 in 2023, then 171.76 in 2024). I do not think this can fairly be called a "remarkable biomass density increase" (Fig. 9 caption). The same applies to the +1.94 % broadleaf change, which is well below the broadleaf class error of about 14 to 18 % in Table 2, and to the +0.44 % total tree-cover change. The manuscript reports static-map accuracy but never quantifies the uncertainty on the change itself. Please add uncertainty or significance estimates for the changes, and adjust the wording ("clear increase", "remarkable") to match what the evidence supports.

Optical saturation is not discussed at all. In closed canopies, optical reflectance saturates above roughly 100 to 150 Mg/ha, so the sensor carries very little information exactly where the Bavarian stocks are highest (coniferous is around 207 Mg/ha in Fig. 9). This is almost certainly the physical reason behind both the modest R2 of 0.62 and the "significant underestimation in the central and mountainous regions" that the authors already note in the Fig. 8 caption. Please discuss saturation explicitly as a physical limit, and explain why 3-D data (GEDI, LiDAR) or SAR (Sentinel-1) were not used, especially since the authors cite exactly that kind of work (Xing et al., 2025; Xu et al., 2022).

There is a resolution mismatch between the biomass reference and the claimed 10 m output. The DLT reference (Copernicus HRL) is 10 m and matches Sentinel-2, but the ESA CCI Biomass product (Santoro and Cartus) has a native resolution of about 100 m. Training and evaluating a 10 m U-Net against a 100 m reference is a tenfold scale gap: the effective resolution of the output cannot really exceed that of the reference, and a per-pixel R2 or rRMSE computed against 100 m labels upsampled to 10 m is hard to interpret. Upsampling those labels also introduces a lot of artificial spatial autocorrelation into the 256 by 256 patches, which tends to inflate the apparent skill. Please state the reference resolution, justify the resampling, and discuss what it means for all the 10 m biomass and degradation results.

I am not convinced the models transfer cleanly across years. They are trained on single-year references (DLT 2018, biomass 2019) but applied from 2015 to 2025. Year-to-year differences in atmosphere, phenology and Sentinel-2 radiometry can easily produce apparent change of exactly the size being reported here. The May to August averaging and the channel-wise normalisation are described, but there is no evidence that the year-to-year artefacts have actually been removed. It would help to show, on the stable-cover reference years, that the detected trends are not just radiometric artefacts.

The conclusion that degradation dominates over deforestation rests on the noisiest layer in the study. Degradation is defined as biomass loss in pixels that show no cover loss, and it is quantified from the biomass model, which has a 45 % relative error. The reported 4.04 to 10.40 % (l.295 to 296) could plausibly sit inside that noise floor, whereas the "fast" deforestation loss is below 0.73 %. Concluding that degradation is the main driver, and that "the impact of logging activities is acceptable" (l.403), is not safe without propagating the error. I would also note that "acceptable" is a policy judgement rather than a scientific result.

The fragmentation trend may be an artefact, and the manuscript half-says so itself. Reason (3) on l.349 already admits that map uncertainty "introduces unreal fragmentation results", and the erosion experiment in Fig. 14 changes the picture substantially, since the 2021 peak disappears. Because the annual maps are predicted independently, the boundary noise varies from year to year and inflates patch and edge density. The strong claim in the abstract that Bavaria's forests are "increasingly fragmented" (l.401) is therefore not robust; I would add a boundary-stability check and soften the statement. Separately, an erosion of 100 pixels is about 1 km (l.356), which seems very large. Please confirm that value and justify it.

The missing 2016 map appears to break the cleaning window. You drop 2016 for lack of data (l.266) but then screen 2018 to 2024 using four consecutive years (the two previous years, the current year, and the next). The window for 2018 needs 2016 to 2019, so it includes the year you removed. Please explain how 2018, and any other window that touches 2016, was actually handled, and why 2016 was discarded rather than gap-filled or interpolated.

The overall accuracy figure is carried by the majority class. The 92.5 % overall is dominated by non-tree at 97 %, while the policy-relevant classes sit at 85 to 87 %. A confusion matrix, and ideally macro-averaged metrics, would be more informative. The conifer-to-broadleaf confusion in mountain areas that the authors flag in the Fig. 5 caption directly affects the species-change and species-attributed biomass results in Sections 5.2, 5.3 and 6.1, so it deserves more than a caption.

There is no independent validation. Both reference datasets are themselves derived from Earth observation, so "92.5 % accuracy" measures agreement with another map, not with ground truth. Calling this "semi-independent" (l.391) is honest, but the German national forest inventory exists for 2012 and 2022 (l.363) and could give at least an aggregate independent check. I would encourage adding one if it is feasible.

The policy link is mostly qualitative, which makes the title a little too strong. Table 1 is a nice framing, but no policy target is actually tested against the data. For instance, the Bavaria 2020 goal of converting 100,000 hectares of spruce to mixed forest (l.110 to 112) is never checked against the derived maps, even though the data could in principle support that. As it stands the paper delivers descriptive trend maps rather than an evaluation of policy outcomes, so I would either add at least one quantitative policy test or tone down the framing.

Reproducibility is limited. As written, another group could not repeat the work. The training is described only by a learning rate of 1e-3 and 50 epochs (Section 4.1); the batch size, weight initialisation, any data augmentation, the stopping criterion, and how the RMSprop settings were chosen are all missing. The tile and patch masking (cloud, no-data, low-biomass) needs enough detail to be repeatable, and the exact CCI Biomass version and access year need to be fixed.

Specific comments
In the abstract (l.8 to 9) and Section 5.2, the statement that the increase "has slowed down since 2020" is only partly right. Total tree cover is roughly flat after 2020 (Fig. 7), but broadleaf keeps rising (16.26 to 17.44 %) and even accelerates in 2023 to 2024. Please make this class-specific.

More generally, the abstract does not quite match the results. It asserts "a clear increase" and the results section calls a 0.4 % change "remarkable", while the body shows changes that are below the model uncertainty (see general point 1). The abstract should summarise what the evidence supports rather than the strongest possible reading, so please align abstract, results and conclusions.

In Section 3.3 (l.167 and l.172) the "midpoint-year" reasoning is inconsistent. 2018 is not the midpoint of 2015 to 2025 (that would be 2020), and 2019 is not the midpoint of 2015 to 2022 or of 2018 to 2024. Please state the real reason, which I suspect is data availability.

In Section 4.1 (l.181 to 182) you say "all the 13 bands are included". Level-2A products carry 12 bands, since the cirrus band B10 is dropped; 13 bands implies Level-1C. Please state which collection and which bands were actually used.

In Eq. (2) (l.201), "linearly normalised via a logarithmic transformation" is self-contradictory. Please give the log base as well. The mask threshold (0.4, said to equal 6.39 Mg/ha) implies a natural log, in which case biomass above roughly 147 Mg/ha maps to values greater than 1, so the claim that most values fall in [0, 1] (l.202 to 203) fails precisely for the high-stock pixels of interest (coniferous is around 207 Mg/ha).

In Eq. (8) the text says edge density is "per km2", but the constant of 10^4 gives metres per hectare, which is the FRAGSTATS convention. Please reconcile text and formula, either using 10^6 for per km2 or stating m/ha.

In Table 2 the test accuracy is higher than the training accuracy for both broadleaf (85.63 vs 82.38) and coniferous (87.35 vs 86.93). That is unusual and should be commented on (spatial split composition, class imbalance), especially since the claim that the model "isn't over-fitted" depends on it.

Fig. 6(b) appears to show an R2 for 2019 of about 0.82, but the test R2 reported in Table 2 (Section 5.3, l.274) is 0.62. Please clarify what each one evaluates (held-out test tiles versus the whole state). Related to this, the sentence spanning pages 15 and 16 (l.282 to 284) states that R2 decreased for 2015 and 2022 while the relative RMSE also decreased, i.e. improved. Since a lower rRMSE indicates a better fit and a lower R2 a worse one, the two moving the same way is counter-intuitive and should be explained (for example through the different denominators: rRMSE is normalised by the mean of the reference values, R2 by their variance, so a year with a narrower biomass distribution can lower R2 without raising rRMSE).

At l.281 (Section 5.3), coniferous biomass is said to be "nearly twice" broadleaf, but 207 versus 129 Mg/ha is about 1.6 times, not nearly twice. Please correct or reword.

Two biomass metrics are used without being clearly separated: "average biomass density" (around 171) and "mean state-level average biomass density" (around 72, l.278 to 279). Please define both precisely (forest-only versus all pixels) so readers can tell which trend is which.

At l.300 (Section 6.1), "significantly larger" is used without any statistical test. Please add a test or drop "significantly".

In Section 6.2 the area of interest is defined by elevation above 1000 m (l.328 to 329) but is repeatedly called "over Alps" (Fig. 12 caption). Much of Bavaria above 1000 m (the Bayerischer Wald, the Fichtelgebirge) is not Alpine. Please rename it to something like "high-elevation regions", or restrict the mask to the Alps.

At l.363 (Section 6.4), the German national forest inventory is not run "every 10 years": 1987 to 2002 is 15 years. Please reword to roughly every 10 to 15 years.

In Section 6.4, point 2 (l.385 to 386), you argue that tree cover should be overestimated relative to the FAO definition, yet 37.35 % is described as "close to" the 37 % inventory figure. If the overestimation were real, your value should sit clearly above 37 %. This tension should be resolved.

The conclusion at l.401 says tree cover "increased by 0.44 % of the total tree area", which contradicts Section 5.2 (l.270 to 271), where it is 0.44 % of the total area. Fig. 7 suggests the denominator is the total state area. Please make results and conclusion consistent.

Technical corrections
l. 215: R2) song2023biomass for quantifying is an unresolved citation key showing up as raw text; it should read (Song et al., 2023).

l. 196: "Senitnel-2" should be "Sentinel-2".

l. 267: "identity suspect predictions" should be "identify".

l. 183: double period ("spatial size..") should be a single period.

l. 146: "forests occupies" should be "forest occupies".

l. 126 (footnote) and the footnotes on p.19 (l.304): bare URLs are pasted into the running text and footnotes without citation dates; please format these as proper references.

Thousands separators are mixed: l.145 has "70,550 km2" with a comma, while l.147 has "26.169 km2" with a period. The period versions are wrong by three orders of magnitude and should read 26,169 km2 (l.147), 2,962 m (l.153) and 14,000 ha (l.326). Please use one convention throughout.

150: "Oak(6.8%)" should be "Oak (6.8 %)".

"(alpha, 2021)" is an awkward citation key for "ARD alpha"; a cleaner key would help.

Fig. 6(a): the inset legend in the top-left overlaps the bars and makes part of the chart unreadable. Please move it outside the plot area or into empty space.

Fig. 13(a) and Fig. 14: putting different-unit metrics on a single rescaled axis ("Patch Density x 10 / Edge Density") is confusing. Twin y-axes would be much clearer.

Fig. 11 is effectively unreadable at print size (seven regions, each with a pie and four bar clusters, in tiny labels). It should be redesigned or split.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-2782-RC1
RC2:
'Comment on egusphere-2026-2782', Anonymous Referee #2, 17 Aug 2026 reply
This manuscript uses Sentinel-2 imagery and U-Net models to generate annual dominant leaf type (DLT) and above-ground biomass (AGB) maps for Bavaria and explores their potential relevance to forest-policy monitoring. The policy-oriented framing is potentially interesting and appropriate for Earth Observation. However, I have substantial concerns about the validation strategy, the spatial scale of the biomass reference data, the interpretation of relatively small temporal changes, and the attribution of biomass loss to degradation, deforestation, and logging. In addition, the connection between the derived EO indicators and actual policy evaluation remains largely conceptual rather than analytical. I therefore recommend major revision. Several of the main conclusions require substantial reanalysis rather than textual clarification alone.
Major comments
Reference products and independent validation. Both models are trained and evaluated primarily against existing remote-sensing products: Copernicus DLT and ESA CCI AGB. Consequently, the reported accuracy values primarily quantify agreement with these products rather than accuracy against independent observations of forest type or biomass. This distinction needs to be made explicit. Independent validation using forest inventory, field, airborne LiDAR, or other suitable data would substantially strengthen the study.

Spatial resolution of the biomass reference requires clarification. The manuscript applies the same 10 m Sentinel-2/U-Net framework to biomass estimation, whereas the ESA CCI Biomass product used as reference has a substantially coarser native resolution. The manuscript does not explain how the reference data were spatially aligned with the 10 m inputs and outputs. If the CCI product was simply resampled to 10 m, the model should be presented as a downscaling or product-emulation approach, and the resulting fine-scale biomass patterns cannot be considered independently validated at 10 m. This is a fundamental methodological issue that needs to be resolved.

Temporal changes need explicit uncertainty assessment. The study reports relatively small changes in state-level tree-cover proportions and biomass density, yet no uncertainty is propagated into these temporal estimates. For example, the reported overall tree-cover increase is approximately 0.44 percentage points, while the DLT model has non-negligible classification error and approximately 4–6% of pixels are excluded during temporal screening. Likewise, the mean biomass density changes only modestly between 2018 and 2024 despite considerable pixel-level prediction error. The authors should quantify uncertainty in the aggregated annual estimates, provide confidence intervals where possible, and demonstrate that the inferred temporal trends are robust to classification and biomass-estimation errors. Expressions such as “remarkable biomass density increase” are currently not supported quantitatively.

The attribution of biomass loss to degradation, deforestation and logging is not supported by the analysis. The manuscript equates biomass decline accompanied by tree-cover loss with deforestation, usually logging, and biomass decline without tree-cover loss with degradation caused by drought, pests, etc. These categories cannot be uniquely inferred from the presented EO products. Harvesting, natural disturbance and deforestation can produce similar canopy-loss signals, while biomass decline without detected canopy loss may arise from several ecological and methodological processes. Therefore, conclusions such as “the impact of logging activities is acceptable” should be removed unless independent disturbance or management datasets are introduced. A more defensible terminology would distinguish “AGB decline associated with detected tree-cover loss” from “AGB decline without detected tree-cover loss”.

The policy contribution should be more clearly defined. Table 1 provides a useful mapping between policy objectives and potentially relevant EO indicators, but the subsequent analyses do not directly evaluate the specified policy interventions. For example, the study does not analyse the spatially explicit areas targeted by the Bavaria 2020 Climate Program or the Mountain Forest Offensive, nor does it evaluate wildlife corridors targeted by connectivity policies. There is also no before-after or treatment-control framework. I therefore suggest either substantially strengthening the spatially explicit policy analysis or reframing the study as the development of policy-relevant monitoring indicators, rather than an evaluation of policy outcomes.

The fragmentation analysis is highly sensitive to methodological choices. The manuscript itself shows that applying a 100-pixel erosion substantially alters temporal trajectories and the ranking of administrative regions. This suggests that the reported fragmentation trends may be strongly influenced by classification uncertainty and unstable forest boundaries. The authors should conduct a systematic multi-scale sensitivity analysis using justified minimum mapping units or boundary treatments and identify which conclusions are robust across reasonable parameter choices.

Temporal transferability requires stronger validation. Models trained using reference data from a single year are applied over approximately a decade. At the same time, the annual Sentinel-2 mosaics are constructed by averaging observations from May to August, and one year is excluded because of extensive missing data. Changes in acquisition timing, cloud availability, phenology and climatic stress may therefore introduce apparent temporal changes unrelated to structural forest change. The manuscript should provide a much more detailed description of annual compositing and cloud masking and validate temporal transferability using all available independent/reference years.

Evaluation metrics and data splitting should be strengthened. For DLT mapping, overall accuracy and class-specific true-positive rates alone are insufficient. A confusion matrix, user’s and producer’s accuracies, precision, recall and preferably F1/IoU should be reported. The spatial train–test split also requires more information on tile size, spatial separation and whether neighbouring training and test tiles may remain spatially autocorrelated. For biomass, bias, MAE/RMSE and residual behaviour across biomass ranges, forest types and geographical regions should be reported in addition to R² and rRMSE.

Dataset versions and reproducibility need correction. The manuscript appears to refer to different ESA CCI Biomass versions in the Methods, references and Data Availability statement, while also evaluating a 2022 reference map. The exact product version, spatial resolution and DOI used for each analysis should be stated consistently. The derived maps and analysis code should also be deposited in a persistent public repository rather than being available only upon request.

Minor comments
AGB and carbon stock are used almost interchangeably throughout the manuscript. AGB in Mg ha⁻¹ is not equivalent to carbon stock unless an explicit biomass-to-carbon conversion is applied.

Changes shown in Figure 7 should be carefully described as percentage-point changes where appropriate. The manuscript is inconsistent in stating that the 0.44% increase refers to total land area or total tree area.

The claim that Bavaria is transitioning towards “mixed forests” is stronger than can be demonstrated from state-level changes in broadleaf and coniferous area. A spatial measure of within-stand or local mixture would be required to support this interpretation.

The definition of the “Alps region” solely as elevation >1000 m should be reconsidered or renamed as high-elevation areas of Bavaria.

The use of all 13 Sentinel-2 bands and nearest-neighbour upsampling of the 20 and 60 m bands to 10 m requires justification.

Statements using “significantly” should be avoided unless supported by statistical testing.

Figure 11 is difficult to interpret because too many quantities are presented simultaneously.

The manuscript requires careful language editing and correction of several formatting/citation issues, including the unresolved “song2023biomass” citation in the evaluation-metrics section.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-2782-RC2

Qian Song and Xiaoshan Ni

Viewed

Total article views: 228 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
140	73	15	228	14	13

HTML: 140
PDF: 73
XML: 15
Total: 228
BibTeX: 14
EndNote: 13

Views and downloads (calculated since 03 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	64	11	9	84
Jul 2026	41	37	1	79
Aug 2026	25	19	5	49
Sep 2026	10	6	0	16

Cumulative views and downloads (calculated since 03 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	64	11	9	84
Jul 2026	41	37	1	79
Aug 2026	25	19	5	49
Sep 2026	10	6	0	16

Viewed (geographical distribution)

Total article views: 189 (including HTML, PDF, and XML) Thereof 189 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 07 Sep 2026

Short summary

We generated the annual dominant leaf type and biomass maps over Bavaria, Germany, from 2015 to 2024. We assessed leaf type and biomass changes across administrative regions and hotspots. The results provide good insights into the tree status in Bavaria and suggest a new focus for forest management policies.


Total:	0
HTML:	0
PDF:	0
XML:	0