Geospatial embeddings detect old-growth forests but buffered spatial validation narrows their advantage over Sentinel features
Thomas Ratsakatika, Mihai Zotta, Srinivasan Keshav, Emily R. Lines
cs.LG · cs.CV · eess.IV
Quick verdict
Maps old-growth forest from free satellite data; TESSERA's +0.08 PR-AUC edge over Sentinel/AlphaEarth loses significance under 10 km buffers; single-site only.
Summary
The paper maps old-growth forest across 211,893 ha of Romania's Făgăraș Mountains using pixel-level XGBoost/CNN classifiers evaluated at forest-parcel resolution. From forest management plans, expert-verified virgin-forest layers, roads and disturbance records, the authors build 1,016 old-growth and 3,837 high-confidence non-old-growth parcel labels, then compare a topographic/access baseline against Sentinel-1/2 composites, AlphaEarth (64-d) and TESSERA v2 (128-d) embeddings under six-fold spatially blocked cross-validation, with and without 10 km train-test buffers sized from measured semivariogram ranges. All EO feature sets beat the baseline by +0.21–0.25 PR-AUC even when buffered. XGBoost on TESSERA reaches PR-AUC 0.84 [0.79–0.88] unbuffered, exceeding Sentinel-1/2 and AlphaEarth by +0.08 each, but this advantage falls to +0.03–0.04 with confidence intervals spanning zero at a 10 km buffer; CNNs add no parcel-level benefit at 10 m resolution. Predicted extent is 55,025 ha [39,124–62,515], with 81% agreement on unlabelled parcels against a field-calibrated map. Limitations: a single landscape with no field validation outside it, reference labels partially sharing provenance with the compared studies, non-probability sampling precluding bias-corrected area estimates, and an unexplained re-emergence of TESSERA's advantage at 20 km buffers.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 0 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Freely available Sentinel-1/2 features add a spatially robust old-growth signal beyond topography/access predictors (+0.21–0.25 PR-AUC at a 10 km buffer, 95% CIs 0.15–0.34).
- TESSERA's unbuffered advantage over Sentinel-1/2 (+0.08 [0.05–0.11]) and AlphaEarth (+0.08 [0.04–0.12]) becomes statistically indistinguishable from zero at a 10 km train-test buffer (+0.04 and +0.03, CIs spanning zero).
- Spatial blocking alone is insufficient for transfer-relevant performance claims; train-test buffering calibrated to measured autocorrelation ranges (10 km here) is required.
- At 10 m resolution, CNN architectures (3×3–7×7 receptive fields) confer no parcel-level benefit over pixel-wise XGBoost (contrasts negative or null across all 20 configurations).
- Explicit high-confidence non-OGF labels materially improve evaluation: existing national/continental maps reach only PR-AUC 0.35–0.65 against these labels, with the field-calibrated rule-based map (0.65) leading model-based products.
Strengths
Spatial-validation protocol is measured, not assumed
Semivariogram ranges (labels 6.6–6.7 km; baseline predictors 16–22 km; §2.6) motivate six blocked folds and a 10 km train-test buffer justified by three independent diagnostics: residual range 4.5–6.5 km (Table S18), performance plateau at 8–10 km against a parcel-matched control (Fig. 5), and the coordinate-only control falling to the 0.21 no-skill floor near 9 km. The control decouples leakage from training-data loss (<0.04 PR-AUC drop at 50% random removal, Figs. S6–S7).
Explicit high-confidence negatives instead of pseudo-absences
§2.2 constructs 3,837 non-OGF parcels (≤80-year stand age, disturbance/road evidence retained) alongside 1,016 OGF parcels, enabling false-positive estimation and PR-AUC — directly addressing the presence-only weakness the authors demonstrate in Sabatini (2020) (PR-AUC 0.35, overlap coefficient 0.80, Fig. S5).
Honest negative results with full uncertainty reporting
Effort-matched, leakage-controlled feature comparison
Reproducibility and data release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether geospatial foundation model (GFM) embeddings improve parcel-level old-growth forest (OGF) detection over conventional Sentinel-1/2 features, and whether any advantage survives spatially rigorous validation. The setting is the Făgăraș Mountains, Romania (211,893 ha; 152,366 ha forest; reference year 2020). Three claims are tested: (1) freely available EO data add old-growth signal beyond topographic/access predictors; (2) AlphaEarth's and TESSERA v2's advantage over Sentinel features depends on validation strictness; (3) CNN spatial context adds nothing at 10 m. A fourth contribution is comparative — four existing maps (Sabatini 2020; Munteanu et al. 2022; Kathmann et al. 2017; Schickhofer & Schwarz 2019) are re-scored against the new labels, their extent estimates varying threefold (20,865–64,146 ha; Table 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 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 old-growth forest detection from free Sentinel satellite data?
At parcel level in this Romanian Carpathian study area, XGBoost with Sentinel-derived features reached PR-AUC 0.77 unbuffered and 0.68–0.73 under stricter 10 km buffered validation, versus a 0.21 no-skill floor at 21% prevalence (best model with TESSERA embeddings: 0.84 unbuffered, 0.73 buffered). ROC-AUC reached 0.91–0.95. Accuracy is highest on steep, high-altitude coniferous terrain and weakest above ~1,400 m.
Do geospatial foundation models (AlphaEarth, TESSERA) beat raw Sentinel features for forest mapping?
Unbuffered, yes: TESSERA beat Sentinel-1/2 and AlphaEarth by +0.08 PR-AUC each (95% CIs excluding zero). But after a 10 km train-test buffer enforcing spatial independence, the advantage fell to +0.03–0.04 with confidence intervals including zero — statistically consistent with no difference. AlphaEarth was statistically indistinguishable from conventional Sentinel features even unbuffered.
How much old-growth forest remains in the Făgăraș Mountains according to this study?
A point estimate of 55,025 ha (95% bootstrap interval 39,124–62,515 ha) out of 152,366 ha of forest — broadly between the four existing national/continental maps (20,865–64,146 ha). The interval reflects resampling of reference labels, not classification error, and the estimate is not bias-corrected. Agreement with the field-calibrated PRIMOFARO map was 80.8% on unlabelled parcels (83.3% overall).
Why did convolutional neural networks (CNNs) not help detect old-growth forests at 10 m resolution?
The authors argue the textural structures signalling old-growth (e.g., canopy gaps — 98% are smaller than 200 m² in Ukrainian Carpathians, about two 10 m pixels) are poorly resolved at Sentinel resolution, and parcel-level aggregation already provides denoising. All CNN receptive fields (3×3 to 7×7, up to 70 m) performed equal to or worse than pixel-wise XGBoost across every feature set.
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.
- Parts of Results §3.2–3.3 and Discussion §4.1–4.2 were omitted in text extraction; conclusions for those spans are reconstructed from remaining prose and tables.
- Figures (1–7, S1–S9) assessed via captions and tables only; map visuals not independently verifiable.
- Supplementary Tables S14–S15 (leave-one-fold-out) are truncated in the extract.
- arXiv v1 preprint dated 23 Sep 2026; peer-review status unknown.
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.