Tree species mapping in Denmark: A comparison of spectral-temporal features with geospatial foundation model embeddings

Data motivation

• Denmark’s stand-level forest dataset is incomplete, while National Forest Inventory plots are sparse and cannot provide wall-to-wall spatial information.The datasets omit privately managed areas or lack the spatial coverage required for operational forest management.Source: §1 · p. 3

Remote-sensing context

• Fine-scale species mapping must preserve temporal dynamics because species discrimination benefits from phenological information beyond single-date or highly aggregated observations.Engineered spectral-temporal features combine Sentinel-2 time-series information with Sentinel-1 structural or textural information.Source: §1 · pp. 3–4

Research gap

• EO foundation-model embeddings offer transferable representations, but it remains unclear whether they can match richer spectral-temporal features in fine-grained tree-species mapping.Prior evaluations largely used coarse-grained benchmarks or simple annual and seasonal composites.Source: §1 · p. 4

Study design

• The study compares Sentinel-1/Sentinel-2 spectral-temporal features with TESSERA and AlphaEarth embeddings, combining each with canopy height and evaluating RF, XGBoost, and MLP classifiers.The framework uses Danish NFI plots as reference data for national-scale dominant tree-species mapping.Source: §1 · pp. 4–5

Evaluation scope

• Performance is assessed separately for pure and mixed stands because mixed stands contain multiple species and present greater structural and spectral complexity.This addresses a gap in evaluations that often focus only on pure stands and provides a more operationally realistic assessment.Source: §1 · p. 5

Contributions

• The study systematically tests training-data availability, spectral indices, Sentinel-1 data, canopy height, and multi-year acquisitions, and releases a 10 m open-access national tree-species map.The map is validated using both pure and mixed tree cover for forest monitoring, biodiversity assessment, and ecosystem management.Source: §1 · p. 5

2.1. Study Area

• Denmark’s forests are small, fragmented, and interspersed with agriculture, urban areas, and infrastructure, creating a challenging context for satellite-based mapping.Forests span varied types and management conditions, making the study area suitable for evaluating machine-learning tree-species classification.Source: §2.1 · p. 6
• Species composition varies regionally with site conditions and management, while mixed-species forests form a substantial share of Denmark’s forests.Broadleaved and coniferous species are nearly evenly represented, and seasonal vegetation dynamics support multi-temporal satellite observations.Source: §2.1 · p. 6

2.2. National Forest Inventory Data

• The NFI uses a nationwide systematic 2 × 2 km grid, with four sample plots clustered at the corners of each 200 × 200 m square and measurements distributed across a five-year cycle.Approximately one-fifth of plots are measured each year across the country.Source: §2.2 · p. 6, Figure 1 · p. 7
• The sampling network contains around 43,000 plots, comprising one-third permanent clusters and two-thirds temporary clusters relocated between five-year cycles.Permanent clusters are remeasured to track forest development, while temporary clusters are randomly relocated within their grid cells.Source: §2.2 · p. 6

2.2.1. Field data collection

• Field protocols geolocate plots precisely, partition them into homogeneous subplots, and record tree species, diameter, selected heights, canopy layer, damage, defoliation, and micro-habitats.Trees are inventoried within nested circular radii of 3.5, 10, and 15 m according to height or diameter thresholds.Source: §2.2.1 · pp. 7–8

2.2.2. Plot selection and species-group labels

• 8,663 plots measured from 2014–2022 remained after removing 225 plots affected by mapped post-measurement disturbances.The initial dataset contained 9,304 unique plots, reduced to 8,888 after removing very rare classes before disturbance filtering.Source: §2.2.2 · pp. 8–9
• The final labels comprise nine species groups, aggregating 59 recorded species to address rarity and spectral similarity in Sentinel observations.The groups are beech, birch, fir, larch, maple, oak, pine, spruce, and other broadleaves.Source: §2.2.2 · p. 9
• 2,683 pure plots and 3,098 mixed plots were retained after filtering plots by forest dominance, stand composition, and suitability for tree-species classification.Pure forest requires domfrac ≥0.90 and standfrac ≥0.80, while mixed forest requires domfrac × standfrac ≥0.50.Source: §2.2.2 · pp. 9–10

2.3. Additional forest layers

• 651,855 ha of rasterized forest extent define the 10 m mapping area, derived from a 2022 Danish Digital Forest Map aligned with the Danish NFI.The original polygon dataset covers approximately 653,250 ha and reports approximately 98% accuracy.Source: §2.3 · p. 10
• A stand-level NST dataset is used for visual assessment, providing dominant tree-species labels for publicly managed forests but omitting many private and commercial forests.Despite incomplete coverage, it supports checks of broad spatial patterns in the resulting map.Source: §2.3 · p. 10

2.4. Sentinel-2

• Sentinel-2 imagery from 2020–2022 undergoes atmospheric, adjacency-effect, bidirectional-reflectance, and topographic corrections, with coregistration, cloud-shadow masking, and 20 m-to-10 m resampling.Processing uses FORCE 3.8.01, the Danish digital elevation model, and Fmask 4.6.Source: §2.4 · p. 10
• Processed imagery and quality data are reprojected to EPSG:25832 and partitioned into non-overlapping 2048 × 2048-pixel tiles at 10 m resolution.The tiles form a national data cube covering Denmark.Source: §2.4 · p. 10
• The data cube and metadata are stored in a STAC, supporting querying, visualization, retrieval, spectral-index calculation, time-series interpolation, and higher-level processing.This infrastructure links the processed imagery to subsequent feature generation.Source: §2.4 · p. 10
• All Sentinel-2 bands except 1, 9, and 10, together with indices targeting canopy phenology, moisture, and chlorophyll properties, are used as classification features.These feature types are selected for their relevance to tree-species discrimination.Source: §2.4 · p. 11
• Spectral bands and indices are smoothed with Whittaker-Eilers interpolation to produce regular 10-day time series and reduce temporal noise.The interpolated series are integrated into the shared data-cube infrastructure.Source: §2.4 · p. 11
• For each year, seasonal summaries include maximum, mean, median, standard deviation, and the 5th, 25th, 75th, and 95th percentiles.These statistics complement the interpolated time series as classification inputs.Source: §2.4 · p. 11

2.5. Sentinel-1

• Sentinel-1 C-band SAR data from 2020–2022 complement Sentinel-2 with forest-structure information from calibrated, filtered, terrain-corrected VV, VH, and VH−VV backscatter time series.Imagery covers Denmark and includes ascending and descending orbits in Interferometric Wide Swath mode.Source: §2.5 · p. 12
• Sentinel-1 observations are aggregated monthly using mean, median, standard deviation, and 25th- and 75th-percentile statistics for VV, VH, and VH−VV.The resulting bands and metadata are integrated into the shared 10 m datacube after reprojection and retiling.Source: §2.5 · p. 12

2.6. Canopy Height Metrics

• Canopy height is estimated by subtracting the 0.4 m Digital Terrain Model from the Digital Surface Model, producing a vegetation-height representation of vertical forest structure.Both elevation components come from the Danish Elevation Model provided by the Danish Climate Data Agency.Source: §2.6 · p. 12
• The 10 m canopy-height inputs comprise spatial median, maximum, standard deviation, and 90th-, 95th-, 98th-, and 99th-percentile metrics.These metrics are added to the satellite-based features for classification models.Source: §2.6 · p. 12
• The STF and foundation-model representations use Sentinel-2 10-day interpolated series and seasonal summaries, monthly Sentinel-1 summaries, annual embeddings, and static canopy-height metrics.These temporal resolutions define the input construction summarized in Table 1.Source: Table 1 · p. 13

2.7. Foundation Model Embeddings

• AlphaEarth [42] and TESSERA provide annual embeddings for 2020–2022 as task inputs, evaluated without fine-tuning or domain-specific adaptation.This tests their out-of-the-box representational capacity in an operational forest-species-mapping setting.Source: §2.7 · p. 14
• The foundation-model inputs consist of dense, analysis-ready representations intended as alternatives to task-specific feature engineering for machine-learning classifiers.The study evaluates pre-trained geospatial models as they become increasingly accessible to the remote-sensing community.Source: §2.7 · p. 14, Table 2 · p. 14

3. Methodology

• The workflow extracts four central 10 m×10 m pixels per NFI plot, trains models on pure plots, evaluates pure and mixed plots separately, and maps Denmark nationally.Pixel predictions are aggregated to plot-level decisions before evaluation and national application.Source: §3 · pp. 14–15, Figure 3 · p. 15

3.1.1. Spectral-temporal features

• The STF representation concatenates Sentinel-2 spectral bands, indices, seasonal statistics, and Sentinel-1 monthly statistics from 2020–2022 with canopy height, producing 3,387 features.The multi-year design accounts for varying cloud contamination, phenology, and climate information across years.Source: §3.1.1 · pp. 15–16

3.1.2. Foundation Model Embeddings

• AlphaEarth and TESSERA embeddings from 2020–2022 are concatenated with canopy height into 199- and 391-dimensional vectors, respectively, without fine-tuning or additional preprocessing.The resulting representations are passed directly to downstream classifiers.Source: §3.1.2 · p. 16

3.2. NFI tree species labeling and splits

• Each plot’s four central pixels inherit its dominant upper-canopy species label, while pure and mixed plots are handled differently because mixed plots introduce greater pixel-label uncertainty.Plot-level labels are assigned to a 2×2 block while avoiding edge-contaminated pixels.Source: §3.2 · p. 16

3.2.1. Training plots and labels

• Classifiers are trained only on pure plots, with plot-level splits assigning 80% to training and 20% to testing while keeping all four pixels from each plot in one split.Within training data, 10% of samples support model optimization and hyperparameter tuning.Source: §3.2.1 · p. 17

3.2.2. Test plots and labels

• Testing combines the remaining 20% of pure plots with a stratified 20% sample of mixed plots, preserving the pre-training pure-to-mixed ratio.During evaluation, a plot is correct when at least two of four pixel predictions match its reference label, reducing effects of sub-plot heterogeneity.Source: §3.2.2 · pp. 17–18

3.3. Classifiers

• RF, XGBoost, and MLP are evaluated under identical conditions and input feature sets, with MLPs using two ReLU hidden layers, AdamW optimization, cross-entropy loss, and 0.5 dropout.Hyperparameters are tuned independently for each classifier and experimental configuration.Source: §3.3 · pp. 18–19

3.4. Validation

• Performance is assessed with per-class and macro-averaged PA, UA, and F1 scores for pure and mixed plots, alongside OA; final-map validation reports area-adjusted metrics and class areas with standard errors.The best-performing model is selected for final map production using the area-estimation and map-accuracy procedure.Source: §3.4 · pp. 19–20

4.1.1. Pure test plots / 4.1.2. Mixed test plots

• The STF MLP achieves the best classification performance on pure forest plots and the highest performance on mixed forest plots, while mixed-plot overall F1 is 19 percentage points lower than on pure plots.Tables 4 and 5 report per-class F1 scores, overall accuracy, macro-averaged F1 score, producer’s accuracy, and user’s accuracy.Source: §4.1.1 · p. 20, Table 4 · p. 20, §4.1.2 · pp. 21–22

4.1.1. Pure test plots

• TESSERA MLP performance was within 1.1 percentage points of the STF MLP for pure stands, whereas AlphaEarth MLP trailed by 10.5pp in F1 score.Across representations, MLP consistently outperformed RF and XGBoost on pure plots.Source: §4.1.1 · pp. 20–21
• STF MLP achieved the highest F1 for most species, while TESSERA MLP performed best for OBL and maple, two underrepresented classes.STF XGBoost performed better for pine and spruce, suggesting pretrained embeddings may partly offset limited class-specific data.Source: §4.1.1 · p. 21

4.1.2. Mixed test plots

• Mixed-stand performance was lower across all models, with the STF MLP losing 19pp in overall F1 relative to pure stands and TESSERA’s gap widening to 5pp.The reduction was consistent across species and classifiers, indicating a challenge associated with heterogeneous stands.Source: §4.1.2 · p. 22
• AlphaEarth consistently underperformed the other representations, so subsequent analysis focused on STF and TESSERA feature sets.This selection followed the mixed-stand comparison, where TESSERA was also less effective at capturing fine distinctions in heterogeneous stands.Source: §4.1.2 · p. 22

4.1.3. Confusion Matrices

• STF and TESSERA MLPs showed largely consistent inter-class confusion patterns across pure and mixed plots, indicating closely aligned discriminatory information.Both models had strong diagonal dominance for pure plots, while mixed plots showed more diffuse errors, including maple–beech and spruce–fir confusion.Source: Table 6 · p. 22, §4.1.3 · pp. 22–23

4.1.4. Limited Training Data

• Below approximately 25% of the original training data, or roughly 500 plots, TESSERA MLP consistently outperformed STF MLP as labels became scarce.STF performance declined more steeply as the training set was reduced from 100% to 1%.Source: §4.1.4 · p. 23

Temporal coverage

• Three-year inputs achieve the highest classification performance for both STF and TESSERA MLPs across pure and mixed forest plots.The comparison covers single-year inputs, consecutive two-year combinations, and the full three-year configuration used in the main experiments.Source: §4.2 · p. 23
• 0.832 is the TESSERA MLP F1 score for the full three-year configuration on pure plots, exceeding its single-year range of 0.771–0.797.Adding a second year produces modest improvements, while the full three-year configuration yields the largest increase.Source: §4.2 · p. 24
• 7.6 pp separates the highest and lowest STF single-year F1 scores on pure plots, compared with a TESSERA range of 0.771–0.797.STF also varies more on mixed plots, ranging from 0.578 to 0.635, while TESSERA ranges from 0.563 to 0.583.Source: §4.2 · p. 24
• TESSERA outperforms STF for 2022 and 2021–2022 inputs, while STF performs better for 2020, 2021, and 2020–2021 inputs.These results indicate that year selection matters and that foundation-model embeddings are more robust to year choice than STF.Source: §4.2 · p. 24

Ablation Study

• 0.843 and 0.653 F1-Scores are achieved by the full feature set for pure and mixed plots, respectively, yielding the highest performance in both forest types.The STF MLP ablation removes each input source from the complete feature set and evaluates pure and mixed forest plots separately.Source: §4.3 · p. 25, Table 8 · p. 25
• 5.7pp for pure plots and 2.0pp for mixed plots are the largest single-source improvements over the Sentinel-2 baseline, produced by Sentinel-1 data.Sentinel-1 is the most influential supplementary component, consistent with sensitivity to canopy structure and cloud-independent observations during spectrally critical phenological periods.Source: §4.3 · pp. 25–26
• Removing canopy height or spectral indices causes modest but consistent performance reductions across both pure and mixed forest categories.These components contribute less than Sentinel-1 but still improve classification when retained in the feature set.Source: §4.3 · pp. 24, 26
• 7.2pp for pure plots and 3.5pp for mixed plots are lost when all three supplementary components are removed, retaining only Sentinel-2 spectral bands.The pronounced degradation indicates that the supplementary inputs are particularly beneficial for forest species mapping.Source: §4.3 · pp. 25–26, Table 8 · p. 25

4.4.1. Visual evaluation

• The STF MLP generated Denmark’s first national 10 m tree species map, providing wall-to-wall predictions beyond existing polygon coverage.The map was produced by applying the best-performing model to Sentinel-1 and Sentinel-2 time series across the country.Source: §4.4.1 · p. 26

4.4.2. Mapped area comparison

• The map’s estimated forest area was 651,855 ha versus 642,380 ha in reference statistics, indicating broadly consistent national area estimates.Pine and OBL agreed within ±3%; oak and birch differed by +13% and −14%, respectively.Source: Figure 7 · p. 28, §4.4.2 · p. 29
• Mapped areas agreed most closely in Nordjylland and Midtjylland, whereas Syddanmark showed the largest regional discrepancies.Many species in Midtjylland were within or near ±10%, while most Syddanmark species deviated by more than 10%.Source: §4.4.2 · p. 29

4.4.3. Map accuracy assessment

• Beech had the highest agreement, with PA 92.4% and UA 85.8%, while pine, fir, and spruce also achieved consistently high accuracies.Map accuracy patterns broadly matched the earlier classification results for dominant coniferous classes.Source: §4.4.3 · pp. 29–30
• 79.9% (±1.3%) overall area-adjusted accuracy was achieved for the national map using combined pure and mixed test plots.The estimate is reported as an unbiased map-level assessment with standard errors.Source: §4.4.3 · p. 29, Table 9 · p. 30
• Area adjustment placed most species within ±10% of reference statistics, reducing direct-map discrepancies by accounting for classification errors.The beech discrepancy decreased by approximately 8,000 ha, and remaining differences were generally within the 95% confidence intervals.Source: §4.4.3 · pp. 30–31, Figure 8 · p. 31

5.1.1. Classification Performance of STF and FM Approaches

• STF remains particularly effective for fine-grained tree-species discrimination, especially in compositionally heterogeneous forests, while shared confusion patterns reflect species similarity and co-occurrence.Larch–spruce percentages are unstable because only 14 pure larch plots were tested; underrepresented classes also show lower accuracy, consistent with scarce training samples and Denmark’s imbalanced species distribution.Source: §5.1.1 · pp. 31–32

5.1.2. Trade-Off Between Accuracy and Operational Demands

• TESSERA’s modest performance gap relative to STF may be acceptable because analysis-ready embeddings reduce preprocessing demands, although they limit input control and interpretability.STF requires radiometric harmonization, cloud masking, co-registration, interpolation, and index calculation, whereas embedding dimensions combine information opaquely across modalities.Source: §5.1.2 · pp. 32–33

5.1.3. Sensitivity to temporal input configuration

• Three-year inputs achieve the highest F1 for both representations, while TESSERA is less sensitive to selected years and outperforms STF for reduced-year pure-plot configurations.For 2022, TESSERA achieves an F1 score of 0.793 versus 0.743 for STF; for 2021–2022, the scores are 0.809 and 0.787, respectively.Source: §5.1.3 · pp. 34–35

5.1.4. Influence of Pre-training Design on Downstream Performance of FMs

• TESSERA consistently outperforms AlphaEarth, indicating that downstream performance depends on alignment between pre-training design and the spatial and thematic granularity of the target task.TESSERA emphasizes pixel-level spectral-temporal dynamics, whereas AlphaEarth integrates broader spatial, temporal, and cross-sensor information, potentially mismatching detailed species discrimination.Source: §5.1.4 · p. 35

5.2.1. Treatment of Mixed Forest

• Mixed-plot performance is lower because dominant-species labels, within-plot heterogeneity, and training exclusively on pure plots create ambiguity and train–test distribution shift.Plot-level aggregation reduces the need for every pixel to be correct, but NFI tree-scaling and pixel-position uncertainties still complicate linking composition to satellite observations.Source: §5.2.1 · pp. 35–36

5.2.2. National-scale mapping and area estimation

• The national map extends continuous dominant-species predictions across Denmark, and area adjustment substantially improves agreement with inventory statistics for most species.For maple, the adjusted estimate differs by 1,071 ha versus 11,741 ha for direct pixel counts; the publicly available map supports operational monitoring and management.Source: §5.2.2 · pp. 36–37

5.3. Limitations and Uncertainties

• Interpretation is limited by upper-canopy sensing, medium-resolution spatial detail, compositional and spatial label noise, non-geographically blocked splits, and untested temporal transfer beyond 2022.These constraints affect both STF and FM embeddings because both derive from the same satellite observations; spatially blocked, cross-region, and future-year validation remain necessary.Source: §5.3 · pp. 37–38

• The STF MLP achieved the highest overall performance, while TESSERA delivered competitive results for pure stands with less preprocessing and no task-specific feature engineering.AlphaEarth embeddings were consistently less effective for this application.Source: §6 · pp. 38–39
• TESSERA embeddings outperformed STF when fewer than 25% of available reference data were used for training, demonstrating greater robustness under limited training data.Source: §6 · p. 39
• Multi-year observations consistently improved classification over single-year inputs, while Sentinel-1 SAR data, canopy height metrics, and spectral indices contributed complementary information.Source: §6 · p. 39
• Mixed forest stands were more difficult to classify than pure stands, emphasizing the importance of accounting for forest compositional complexity in operational mapping.Source: §6 · p. 39
• 79.9% overall accuracy was achieved by the first open-access 10 m national tree species map of Denmark, produced using the best-performing STF MLP.Area-adjusted validation indicated that the map reliably represented dominant tree species patterns at national scale.Source: §6 · p. 39

• Alkiviadis Koukos led original-draft writing and contributed to conceptualization, methodology, software, validation, formal analysis, investigation, data curation, and visualization; Spyros Kondylatos contributed across review and editing, research design, analysis, investigation, validation, software, and visualization.Source: CRediT authorship contribution statement · p. 39
• Thomas Nord-Larsen contributed writing review and editing, resources, validation, and data curation; Lotte Nyborg contributed validation, resources, funding acquisition, and project administration; Christian Tøttrup contributed funding acquisition and project administration; Kenneth Grogan contributed review and editing, conceptualization, methodology, validation, formal analysis, investigation, visualization, and supervision.Source: CRediT authorship contribution statement · pp. 39–40