Tree species mapping in Denmark: A comparison of spectral-temporal features with geospatial foundation model embeddings
Alkiviadis Koukos, Spyros Kondylatos, Thomas Nord-Larsen, Lotte Nyborg, Christian Tøttrup, Kenneth Grogan
cs.CV · cs.AI · cs.LG
Quick verdict
Careful national mapping study: STF beats FM embeddings at full data (pure-stand F1 0.843 vs 0.832); TESSERA wins below ~25% training data; minor stat gaps.
Summary
The paper presents a national-scale framework for mapping dominant tree species across Denmark at 10 m resolution, comparing manually engineered spectral-temporal features (STF) from 2020–2022 Sentinel-1/2 data (3,387 features) against out-of-the-box embeddings from the TESSERA (391-d) and AlphaEarth (199-d) foundation models, both augmented with seven canopy-height metrics. Using 8,663 Danish NFI plots (2,683 pure for training/testing; 3,098 mixed for testing only), the authors train Random Forest, XGBoost, and MLP classifiers under a plot-level 80/20 split, evaluating at plot level via majority vote of four overlapping pixels. The STF MLP performs best, with macro F1 of 0.843 (pure) and 0.653 (mixed); the TESSERA MLP trails by only 1.1 pp on pure stands but 5 pp on mixed stands, while AlphaEarth lags by 10.5 pp. TESSERA outperforms STF when fewer than ~25% of training plots are used and is less sensitive to which single year is used. Ablations identify Sentinel-1 backscatter as the most influential addition; multi-year inputs consistently help. The best model yields a wall-to-wall map with area-adjusted overall accuracy of 79.9% (±1.3), released open-access. Main limitations: a single random train/test split without variance estimates or significance tests for model comparisons, narrative numbers in the ablation section that do not match Table 8, no mixed-stand training baseline, and acknowledged lack of spatially blocked validation.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Engineered STF remain the best representation at full training data: the STF MLP reaches macro F1 0.843 (pure) and 0.653 (mixed), ahead of the TESSERA MLP (0.832 / 0.603) and far ahead of AlphaEarth (10.5 pp behind on pure).
- TESSERA embeddings outperform STF when training plots fall below ~25% (~500 plots), with the advantage widening as data shrinks (Figure 5).
- TESSERA is more robust than STF to year selection (2022-only: 0.793 vs 0.743 macro F1 on pure plots, Table 7), and three-year inputs improve every configuration.
- Sentinel-1 backscatter is the most influential supplementary input; removing it from the full STF set costs 5.0 pp macro F1 on pure plots (Table 8).
- The resulting 10 m national map achieves area-adjusted overall accuracy of 79.9% (±1.3) and is the first open-access high-resolution tree species map of Denmark.
Strengths
Design-based map validation with uncertainty
Table 9 applies Olofsson et al. [71] via the mapac package: area-adjusted overall accuracy 79.9% (±1.3) with per-class PA/UA/F1 and standard errors — rare rigor in national species mapping.
Fair, controlled comparison design
STF and FM representations receive identical canopy-height complements, the same three classifiers, per-configuration hyperparameter tuning, and plot-level splits that keep each NFI plot's four pixels in one partition (§3.2, §3.3).
Realistic evaluation including mixed stands
Concrete open deliverable and honest area accounting
Informative ablations and low-data analysis
Weaknesses
No uncertainty for model comparisons
Ablation narrative conflicts with Table 8
Temporal label/feature mismatch under-controlled
No mixed-training baseline or spatially blocked validation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv:2609.03480v3) asks whether analysis-ready embeddings from Earth-observation foundation models (TESSERA, AlphaEarth) can replace manually engineered spectral-temporal features (STF) for national-scale dominant tree species mapping. Using the Danish NFI (8,663 plots after filtering; 2,683 pure for training/test, 3,098 mixed for test only), Sentinel-1/2 data from 2020–2022, and canopy-height metrics, the authors train RF, XGBoost, and MLP classifiers and release the first open 10 m national species map. Central claims: (1) STF-MLP is best (macro F1 0.843 pure / 0.653 mixed, Tables 4–5); (2) TESSERA-MLP is within 1.1 pp on pure stands and wins below ~25% training data (Fig. 5); (3) TESSERA is more robust to year choice (2022-only: 0.793 vs 0.743, Table 7); (4) area-adjusted map accuracy is 79.9% ±1.3 (Table 9).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 5 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 accurate is the new Denmark tree species map?
Area-adjusted validation gives 79.9% (±1.3) overall accuracy across nine species groups at 10 m resolution. Beech is most reliable (F1 89.0%), maple least (F1 64.4%, producer's accuracy 51.9%). Accuracy refers to dominant species within the mapped forest extent, which itself comes from a Danish forest mask with ~98% reported accuracy.
Do foundation model embeddings beat handcrafted Sentinel features for tree species classification?
Not when full training data are available: the STF MLP reaches macro F1 0.843 vs 0.832 for TESSERA on pure plots (and 0.653 vs 0.603 on mixed). TESSERA wins when training plots drop below ~25% and tolerates single-year inputs better; AlphaEarth trails the best model by 10.5 pp on pure stands.
Where can I download the Denmark tree species map?
The authors released the 10 m national map openly on Zenodo (record 22108850). The paper does not state that training code is released; preprocessing used public tools (FORCE, Fmask, SNAP, mapac).
Why is mixed-forest classification accuracy so much lower?
Models are trained only on pure plots, and a single dominant-species label cannot describe heterogeneous pixels, so macro F1 falls from 0.843 (pure) to 0.653 (mixed). The authors attribute this jointly to reference-label ambiguity, within-plot heterogeneity, and train-test distribution shift.
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.
- Numerical values plotted in Figures 5, 7, and 8 (limited-data curves, per-region areas) exist only as graphics; the review relies on the authors' in-text descriptions of them.
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.