National-scale mapping of tree species composition in Switzerland from high-resolution aerial imagery and multimodal data
Authors: Yixin Zhou, Luca Ferrari, Verena C. Griess, Nataliia Rehush, Konrad Schindler, Mirela Beloiu
Organizations: Photogrammetry and Remote Sensing Group, ETH Zurich, Zurich, 8093, Switzerland · Forest Resources Management Group, ETH Zurich, Zurich, 8093, Switzerland · Ecosystem Ecology Group, Swiss Federal Research Institute WSL, Zürcherstrasse 111, Birmensdorf, 8902, Switzerland · ETH Swiss Geo Lab, D4 Business Village, Platz 1a, Root, 6039, Switzerland · Forest Resources and Management, Swiss Federal Institute for Forest, Snow and Landscape Research WSL, Birmensdorf, 8903, Switzerland
Tree-species composition shapes forest productivity and how forests respond to climate disturbances such as drought, windthrow and insect outbreaks. Yet most national forest maps assign each pixel a single dominant species and cannot show the share of accompanying species. We mapped the basal-area shares of ten main tree species and a residual Other class across Switzerland at 30 m resolution. A deep-learning model combining 0.1 m aerial RGB imagery, airborne laser scanning (ALS), annual AlphaEarth satellite embeddings, and climate variables was trained on basal-area observations from 3,992 Swiss National Forest Inventory (NFI) sites. In spatial cross-validation, the model reached a macro-R2 of 0.485 across the 11 classes. On an independent forest-reserve dataset, the map reached a macro-R2 of 0.552 for individual 30 m pixels and 0.733 for plots, and identified the dominant class in 73.2% of pixels and 81.1% of plots. The map shows maple and ash as widespread minor components that dominant-class maps omit. It also reproduces the elevational sequence from broadleaved lowland stands through spruce to larch and Arolla pine near the treeline. By resolving the species mixture within each 30 m pixel, the map complements field inventory statistics with spatially continuous information that can support forest monitoring, the management of mixed stands, and climate-adaptation planning across Switzerland.
Figures & tables
Study/product
Prediction target
Predictor modalities
Scale
Categorical
Francini et al. (2024)
7 species groups
S-2
Netherlands
Mu et al. (2025a)
8 species/species groups
S-2
Germany
Waser et al. (2017)
2 leaf types; broadleaf probability
Aerial multispectral + terrain
Switzerland
Blickensdörfer et al. (2024)
11 species/species groups
S-1 + S-2 + environment
Germany
Koch et al. (2024)
15 species
S-2 + terrain
Switzerland
Table 1: Large-area studies and products mapping tree species or leaf type. S-1/S-2 = Sentinel-1/2; ALS = airborne laser scanning; HSI = hyperspectral imagery; AEF = AlphaEarth Foundations satellite embeddings.
Figure 1: Overview of the Swiss National Forest Inventory (NFI) sites used in our study. (a) Number of retained NFI sites in each of Switzerland’s 26 cantons, in five classes (3,992 sites in total); cantons are labeled with their official two-letter abbreviations. (b) Number of retained sites by NFI survey year (2017–2022). (c) Number of retained tree-level observations assigned to each of the 10 Swiss NFI main tree species and the residual Other class, colored by leaf type. Here and in all other maps of this article and its Supplementary Material, map lines delineate study areas and do not necessarily depict accepted national boundaries.
Modality
Dataset
Spatial resolution
Acquisition time
Aerial RGB
SWISSIMAGE
0.1 m
Leaf-on acquisition nearest the NFI survey year
ALS
swissSURFACE3D
15–20 points m -2
Single available acquisition
Satellite embeddings
AlphaEarth Foundations
10 m
NFI survey year
Climate
CHELSA BIOCLIM+
∼ 1 km
1981–2010 long-term mean
Table 2: Input data modalities. ALS = airborne laser scanning.
Figure 2: Forest reserve validation data. Left: locations of the 132 plots of the Swiss Forest Reserves Monitoring network used for independent validation. Right: an example plot polygon (red outline), with circles marking the retained trees colored by mapped class and sized by DBH.
Figure 3: Workflow of this study. (a) Construction of site-level basal-area compositions from NFI tree records. (b) Extraction and screening of the four input modalities. (c) Model training and five-fold spatial block cross-validation. (d) National mapping with the refitted model and independent evaluation.
Figure 4: Overview of the multimodal architecture. Modality-specific encoders project each input into a token set in a shared embedding space: a frozen DINOv3 ViT-H+/16 for the SWISSIMAGE crop, a sparse 3D U-Net with learned queries for the ALS point cloud, and multilayer perceptrons for the AlphaEarth patch and the CHELSA variables (Section 3.1.1 ). A learned query pools the aerial features into the anchor h0 ; the other modalities enter as residual corrections Δm , each scaled by a learned gate gm (Section 3.1.2 ). Three heads predict stand type, leaf type, and species composition, with the loss terms of Section 3.1.3 .
Figure 5: Spatial cross-validation design and example reference sites. (a) The 3,992 retained NFI sites (gray dots) were grouped into 10 km spatial blocks, and each block was assigned to one of the five cross-validation folds; the bar chart gives the mean site-level basal-area share of each mapped class in the five folds. Numbered markers (1–10) locate the example sites in (b). (b) One pure 30 × 30 m site for each of the ten main tree species, with white circles marking the recorded NFI trees and the frame color indicating leaf type.
Model
Backbone
Leaf level
Species level
MAE
R2
MAE
R2
F1
ResNet-50
ResNet-50, ImageNet
0.109 ± 0.004
0.757 ± 0.020
0.058 ± 0.002
0.380 ± 0.043
0.515 ± 0.027
Random Forest
–
0.167 ± 0.004
0.669 ± 0.017
0.083 ± 0.001
0.313 ± 0.012
0.342 ± 0.031
AlphaEarth MLP
–
0.149 ± 0.008
0.667 ± 0.027
0.084 ± 0.002
0.157 ± 0.014
0.194 ± 0.012
Ours
ViT-L/16, SAT-493M
0.103 ± 0.006
0.781 ± 0.020
0.052 ± 0.002
0.446 ± 0.062
0.570 ± 0.043
ViT-L/16, LVD-1689M
0.097 ± 0.004
0.794 ± 0.013
0.051 ± 0.002
0.444 ± 0.063
0.566 ± 0.051
Table 3: Comparison with reference baselines and across frozen DINOv3 backbones. Best value in bold, second best underlined.
Modality
Leaf level
Species level
RGB
ALS
AEF
CHELSA
MAE
R2
MAE
R2
F1
✓
✓
✓
✓
0.093 ± 0.003
0.803 ± 0.017
0.048 ± 0.001
0.485 ± 0.038
0.585 ± 0.014
✓
✓
✓
0.094 ± 0.002
0.808 ± 0.013
0.049 ± 0.002
0.465 ± 0.056
0.581 ± 0.031
✓
✓
✓
0.093 ± 0.002
0.814 ± 0.013
0.048 ± 0.001
0.483 ± 0.052
0.593 ± 0.026
✓
✓
✓
0.099 ± 0.004
0.791 ± 0.017
0.050 ± 0.002
0.464 ± 0.070
0.580 ± 0.029
✓
✓
0.099 ± 0.003
0.793 ± 0.018
0.051 ± 0.001
0.461 ± 0.044
0.581 ± 0.027
Table 4: Leave-one-modality-out and reduced-modality ablations. AEF = AlphaEarth annual embeddings. Best value in bold, second best underlined.
Figure 6: Dominant mapped class of the national product. Top: the class with the largest predicted basal-area share. For display, the 30 m map was aggregated to 210 m by majority class, and cells with less than 50% forest cover are left blank; rings give the composition of three cantons. Minor classes are shown in Fig. 7 . Bottom: three forest reserves of Section 4.3.1 , marked a–c above, at the native 30 m resolution. Red outlines are the reserve plot polygons.
Figure 7: Predicted basal-area composition of each mapped class across Switzerland. Every panel gives the predicted basal-area proportion of one class on a common linear 0–1 scale: a–e the five coniferous groups, f–j the five broadleaved groups, and k the residual Other class. For display, the 30 m predictions were averaged over the forest pixels of 990 m cells, and only cells with at least 25% forest cover are shown. Gray marks forest in which the proportion of the class is below 0.005; areas without forest are white.
Metric
Spatial 5-fold CV
Independent, plot-level
Independent, pixel-level
NFI ( n=3,992 )
Forest reserve ( n=132 )
Forest reserve ( n=414 )
R2
0.483
0.733
0.552
MAE
0.048
0.038
0.042
RMSE
0.125
0.088
0.109
Dominant-class accuracy
0.734
0.811
0.732
Dominant-class F1
0.585
0.704
0.596
Table 5: Accuracy across evaluation supports. R2 is averaged over the ten classes other than chestnut, and MAE and RMSE over all 11 classes.
Figure 8: Observed and predicted basal-area proportion of each mapped class. (a) At the 132 forest reserve plots (plot-level support). (b) At 414 forest-reserve pixels, reconstructed on the product’s native 30 m grid (pixel-level support). The dashed line is the 1:1 relationship. Panel annotations give the per-class MAE, R2 , and the number of plots or pixels in which the class was observed ( n+ ). n/a: R2 not evaluable.
Figure 9: Row-normalized confusion matrix of the dominant class at (a) the 132 forest reserve plots (plot-level support) and (b) the same 414 forest-reserve pixels (pixel-level support). Cell labels give the percentage of the observed row total; n in the row labels is the number of plots or pixels in which each class was the observed dominant one. Chestnut never dominated a plot or pixel and is therefore not evaluable as an observed class.
Figure 10: Agreement between the national product and NFI4 ( Abegg et al., 2023 ) across the five Swiss production regions (Fig. S5 ). (a) Basal-area composition of each region in the product and in NFI4; segments below 5% are left unlabeled. (b) The same 55 region-by-class pairs, with color denoting the mapped class and marker shape the production region. The inset enlarges the range below 10%, in which most pairs lie; the dashed line is the 1:1 relationship and the solid line the ordinary least-squares fit, and MAE and RMSE are taken about the 1:1 line in percentage points (pp).
Figure 11: Predicted tree-species composition along the Swiss elevation gradient. (a) The five coniferous groups. (b) The five broadleaved groups and the residual Other class, sharing the elevation axis with (a). Curves give the mean predicted basal-area share of each class in each elevation band, with elevation from swissALTIRegio. The gray backdrop, read against the right-hand axis in (b), gives the mapped forest area of each band. Only bands containing at least 5,000 forest pixels (4.5 km 2 ) are shown.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Step
Screening criterion
Description
Tree points
Sites
1
Raw NFI tree observations
All available tree records
63,969
5,998
2
Recent inventory observations
Recorded in 2017 or later
42,831
4,052
3
Canopy-layer screening
Exclude the lowest stratum
39,863
4,043
4
Quality control and completeness
Exclude disturbed sites and sites with degraded imagery; require complete multimodal records
39,380
3,992
Appendix
Table S1: Attrition of tree-level observations and sites through the eligibility filtering pipeline.
Figure S1: Conceptual examples of the three principal site-level basal-area distribution forms and their complementary targets. The examples represent a pure stand, a mixed stand within one leaf type, and a mixed stand spanning both leaf types. Left: aerial imagery and NFI sample circles; center: schematic class composition and the 0.80 stand-type threshold; right: the corresponding broadleaved–coniferous composition.
Figure S2: GPS positional extent of the retained NFI tree records at 3,992 eligible sites. Left: cumulative distribution of the minimum enclosing-circle radius of retained tree positions; vertical dashed lines mark 5 m and 10 m. Right: containment within centered square pixels of increasing side length; the solid curve gives sites fully contained, the dashed curve individual tree points contained, and the vertical dashed line marks the 30 m national-product grid.
Pixel side length
10 m
14 m
18 m
20 m
24 m
30 m (used here)
Sites fully contained (%)
10.9
24.0
49.2
65.5
90.6
99.0
Tree points contained (%)
34.6
61.2
82.7
90.2
98.1
99.9
Appendix
Table S2: Containment of retained NFI tree positions within a square output pixel of the given side length, centered on the site coordinate, for the 3,992 eligible sites.
Figure S3: Reference year of the national-scale inference. Each forest pixel was assigned the year of the most recent leaf-on SWISSIMAGE acquisition available for its area, and that year also determined which annual AlphaEarth embedding was used. Percentages give the share of the mapped forest area falling in each year.
Table S3: Definition of the 11 mapped classes. The ten main tree species follow the Swiss NFI grouping; Other contains all retained taxa outside those ten. English and Latin give the common and scientific name of each class, respectively.
Taxon
Leaf type
Taxon
Leaf type
Abies alba
Coniferous
Pinus sylvestris
Coniferous
Acer campestre
Broadleaved
Platanus spec.
Broadleaved
Acer opalus
Broadleaved
Populus alba
Broadleaved
Acer platanoides
Broadleaved
Populus nigra s.l.
Broadleaved
Acer pseudoplatanus
Broadleaved
Populus spec.
Broadleaved
Aesculus hippocastanum
Broadleaved
Populus tremula
Broadleaved
Appendix
Table S4: Assignment of the 71 retained NFI taxa to broadleaved or coniferous leaf type.
Figure S4: Seasonal NDVI trajectories by mapped class, derived from the quality-screened Sentinel-2 time series. Lines show the class median across sites; shaded bands show the bootstrap confidence interval of the median.
Variable
Interpretation
Unit
bio01
Annual mean temperature
∘ C
bio02
Mean diurnal temperature range
∘ C
bio03
Isothermality (bio02/bio07 × 100)
%
bio04
Temperature seasonality (standard deviation × 100)
∘ C × 100
bio05
Maximum temperature of the warmest month
∘ C
bio06
Minimum temperature of the coldest month
∘ C
Appendix
Table S5: Definitions of the 19 CHELSA BIOCLIM+ variables used by the model.
Modality
Model input
Encoder
Trainable
Aerial imagery
280 × 280 × 3 crop, resized to 256 × 256 (28 m, 0.1 m GSD)
Figure S5: The five Swiss NFI production regions used for the comparison with NFI4 regional statistics (Section 4.3.2 ).
Variant
Leaf level
Species level
MAE
R2
MAE
R2
F1
AlphaEarth embeddings †
0.093 ± 0.003
0.803 ± 0.017
0.048 ± 0.001
0.485 ± 0.038
0.585 ± 0.014
Sentinel-2, July composite
0.098 ± 0.004
0.791 ± 0.015
0.050 ± 0.001
0.464 ± 0.044
0.588 ± 0.029
Sentinel-2, growing season
0.099 ± 0.003
0.786 ± 0.018
0.050 ± 0.001
0.458 ± 0.052
0.579 ± 0.041
Sentinel-2, all year
0.098 ± 0.003
0.790 ± 0.011
0.050 ± 0.001
0.480 ± 0.042
0.593 ± 0.020
Appendix
Table S8: Satellite representation. Best in bold, second best underlined. The dagger marks the configuration retained in the reference model.
Variant
Leaf level
Species level
MAE
R2
MAE
R2
F1
CHELSA BIOCLIM+ †
0.093 ± 0.003
0.803 ± 0.017
0.048 ± 0.001
0.485 ± 0.038
0.585 ± 0.014
Climplicit embedding
0.092 ± 0.003
0.813 ± 0.015
0.049 ± 0.001
0.487 ± 0.042
0.593 ± 0.027
Location MLP
0.093 ± 0.003
0.815 ± 0.014
0.049 ± 0.002
0.473 ± 0.044
0.585 ± 0.023
Appendix
Table S9: Climate representation. Best in bold, second best underlined. The dagger marks the configuration retained in the reference model.
Figure S6: Sensitivity of the pixel-level validation to the minimum-tree-count threshold, with pixel coverage fixed at 80%. The dashed line marks the primary specification (10 trees) used in Section 4.3.1 .
Class set
Spatial 5-fold CV
Independent, plot
Independent, pixel
NFI ( n=3,992 )
Forest reserve ( n=132 )
Forest reserve ( n=414 )
All 11 classes (including chestnut)
0.485
−3.560 a
n.e. b
Ten evaluable classes (excluding chestnut)
0.483
0.733
0.552
Appendix
Table S10: Sensitivity of macro-averaged R2 to the inclusion of chestnut across evaluation supports. Cross-validation values are means of the five fold-level macro- R2 values. a At plot support, the value including chestnut is reported only as a diagnostic because chestnut has a single non-zero reference observation. b At pixel support, chestnut has zero reference variance and R2 is therefore not evaluable (n.e.). The second row uses the same ten evaluable classes at all three supports.
Class
MAE
RMSE
R2
n+
Plot
Pixel
Plot
Pixel
Plot
Pixel
Plot
Pixel
Spruce
0.085
0.115
0.172
0.237
0.668
0.677
89
264
Fir
0.034
0.038
0.071
0.092
0.829
0.643
48
100
Pine
0.044
0.063
0.135
0.194
0.718
0.421
50
102
Larch
0.020
0.029
0.058
0.085
0.869
0.728
23
61
Arolla pine
0.031
0.039
0.118
0.138
0.700
0.610
18
56
Appendix
Table S11: Per-class performance at the plot-level (132 forest reserve plots) and pixel-level (414 forest-reserve pixels) independent validations. n+ is the number of plots or pixels in which the class was observed with a non-zero share. Chestnut has a single non-zero plot-level observation and no positive pixel-level observation, so its R2 is not evaluable (n.e.) at either support and it is excluded from the macro- R2 ; the macro-MAE and macro-RMSE retain all 11 classes. Dominant-class support at each observation unit is given in the row labels of Fig. 9 .
Class
MAE
R2
CV
Pixel-level
CV
Pixel-level
Spruce
0.111
0.115
0.746
0.677
Fir
0.061
0.038
0.668
0.643
Pine
0.024
0.063
0.599
0.421
Larch
0.039
0.029
0.742
0.728
Arolla pine
0.013
0.039
0.339
0.610
Appendix
Table S12: Per-class MAE and R2 in the spatial five-fold cross-validation and the pixel-level independent validation. Cross-validation values are the mean over the five folds, restricted to the reference configuration. Chestnut is omitted from this side-by-side table because its pixel-level R2 is not evaluable.
We map tree species across Denmark using National Forest Inventory plots and EO data, while evaluating the potential of foundation models for large-scale forest characterization. We compare two alternative input representations for tree species classification: (i) manually engineered spectral-temporal features (STF) derived from multi-temporal Sentinel-1 and Sentinel-2 observations, and (ii) embeddings generated by the EO FMs TESSERA and AlphaEarth. Both representations are complemented with canopy height information. Random forest, XGBoost, and Multi-Layer Perceptron (MLP) classifiers are evaluated for all input representations, with separate assessments for pure and mixed forest stands. The STF-based MLP achieves the highest classification performance, yielding macro F1 scores of 0.843 and 0.653 for pure and mixed stands, respectively. The MLP trained on TESSERA embeddings delivers competitive performance for pure stands, achieving results within 1.1 percentage points of the best-performing model. TESSERA consistently outperforms STF-based models when fewer than approximately 25% of training plots are available, demonstrating a substantial advantage under limited training data. Multi-year observations systematically improve classification accuracy relative to single-year inputs, while ablation experiments reveal the complementary contributions of Sentinel-1 backscatter, spectral indices, and canopy height data. The best-performing model is subsequently applied at the national scale to generate a 10 m tree species map of Denmark. Area-adjusted validation indicates an overall map accuracy of 79.9%. The resulting map, released as an open-access product, is the first high-resolution national tree species map of Denmark and provides a valuable resource for forest monitoring, ecological research, and land management applications.
Alkiviadis Koukos, Spyros Kondylatos, Thomas Nord-Larsen +3
EO Centre of Excellence, DHI, Agern Alle 5, Hørsholm, 2970, Denmark · Department of Geoscience and Natural Resource Management, University of Copenhagen, Øster Voldgade 10, Copenhagen, 1350, Denmark · ceometrics, Strandvejen 273C, Charlottenlund, 2920, Denmark
The shift from stand-level to individual-tree-level forest assessments supports improved species mapping and biodiversity monitoring, particularly in boreal ecosystems where tree species like aspen (Populus tremula L.) play a keystone role. Airborne laser scanning (ALS) is the standard for such inventories, but a major limitation for developing improved species classification methods is the small number of publicly available ALS datasets containing high-quality, field-validated reference data. Recently, multispectral ALS data has shown promise for tree species classification, but the progress is hindered by the lack of open multispectral ALS datasets with high-quality field reference data. This paper presents and details an open multispectral ALS dataset for tree species classification that was used before its public release for an international benchmarking study of machine learning and deep learning classification methods in a related publication by Taher et al.,(2026). The dataset comprises 6326 segment-level point clouds of individual trees representing nine species in southern Finland. The point cloud data has been acquired using two multispectral laser scanning systems each operating at three laser wavelengths: a helicopter-borne system (HeliALS) with a point density exceeding 1000 points\m2 and an Optech Titan system with approximately 35 points\m2. Furthermore, we present a crowdsourcing application that facilitates the collection of high-quality field reference data of tree species in an efficient and scalable manner. Our article showcases the versatility of the open dataset by presenting new analyses on species classification using multispectral data building upon the initial findings of Taher et al.,(2026).
Matti Hyyppä, Klaara Salolahti, Eric Hyyppä +9
Department of Remote Sensing and Photogrammetry, Finnish Geospatial Research Institute FGI, The National Land Survey of Finland, Vuorimiehentie 5, Espoo, FI-02150, Finland
Trees Outside Forests (TOF) play an important role in agricultural landscapes by supporting biodiversity, sequestering carbon, and regulating microclimates. Yet, most studies have treated TOF as a single class or relied on rigid rule-based thresholds, limiting ecological interpretation and adaptability across regions. To address this, we evaluate deep learning for TOF classification using a newly generated dataset and high-resolution aerial imagery from four agricultural landscapes in Germany. Specifically, we compare convolutional neural networks (CNNs), vision transformers, and hybrid CNN-transformer models across six semantic segmentation architectures (ABCNet, LSKNet, FT-UNetFormer, DC-Swin, BANet, and U-Net) to map four categories of woody vegetation: Forest, Patch, Linear, and Tree, derived from previous studies and governmental products. Overall, the models achieved good classification accuracy across the four landscapes, with the FT-UNetFormer performing best (mean Intersection-over-Union 0.74; mean F1 score 0.84), underscoring the importance of spatial context understanding in TOF mapping and classification. Our results show good results for Forest and Linear class and reveal challenges particularly in classifying complex structures with high edge density, notably the Patch and Tree class. Our generalization experiments highlight the need for regionally diverse training data to ensure reliable large-scale mapping. The dataset and code are openly available at https://github.com/Moerizzy/TOFMapper
Moritz Lucas, Hamid Ebrahimy, Viacheslav Barkov +3
Osnabrück University, Institute for Computer Science, Remote Sensing Osnabrück, Germany · Leibniz Institute for Agricultural Engineering and Bioeconomy, System Process Engineering, Germany · Osnabrück University, Institute of Cognitive Science, Artificial Intelligence, Germany +1