PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.28194v1cs.LGcs.CVeess.IV

TL;DR

Old-growth mapping must be accurate at parcel scale yet deployable continent-wide, but geospatial foundation models had not been evaluated for this task. The study compares EO feature sets with baseline predictors using spatially blocked validation, with and without train-test buffers; TESSERA led without buffering, while feature sets were consistent with no difference with 10 km buffers.

  • Problem

    Old-growth maps must be accurate at parcel scale yet deployable continent-wide, while geospatial foundation model embeddings had not been evaluated for this detection task.

    Source: §1 · pp. 1–2

  • Method

    The study compares Sentinel-1/2, AlphaEarth, and TESSERA v2 features against topographic and access predictors using high-confidence labels and spatial validation with buffering based on measured autocorrelation.

    Source: §1 · p. 2

  • Results

    PR-AUC 0.84 [95% CI 0.79–0.88] was achieved within the AOI; TESSERA led without buffering, but EO feature sets were consistent with no difference with a 10 km buffer.

    Source: §5 · p. 14

  • Takeaways & Limitations

    Spatial blocking alone was insufficient to control spatial autocorrelation; train-test buffering is required when evaluating transfer to unseen landscapes.

    Source: §5 · p. 14

    The study covers one mountain beech–spruce landscape, and accuracy was not tested outside the AOI or verified in the field.

    Source: §4.6 · p. 13

Abstract

from arXiv · show

Old-growth forests develop over centuries under minimal anthropogenic disturbance, producing structurally complex and biodiverse stands. In Europe, protecting them requires mapping that is accurate for individual forest parcels yet deployable continent-wide. Geospatial foundation model (GFM) embeddings enable label-scarce land classification, but their value for old-growth detection remains unknown. Here, we map old-growth forests across 211,893 ha of Romania's Southern Carpathians, a beech-spruce landscape typical of the Alpine Biogeographic Region. We construct high-confidence, expert-informed reference labels for old-growth and non-old-growth parcels. We add AlphaEarth, TESSERA v2 and Sentinel-1/2 features to a common baseline of topographic and human-access predictors, then compare them under spatially blocked validation with and without 10 km train-test buffers to limit residual autocorrelation. With buffering, GFM and Sentinel-1/2 predictors increase precision-recall AUC by 0.21-0.25 [95% CIs: 0.15-0.34] relative to baseline, indicating spectral data contain a spatially robust old-growth signal. With a PR-AUC of 0.84 [0.79-0.88], TESSERA outperforms Sentinel-1/2 (+0.08 [+0.05 to +0.11]) and AlphaEarth (+0.08 [+0.04 to +0.12]) under unbuffered spatial validation. At a 10 km buffer, however, this advantage narrows to +0.04 [-0.01 to +0.11] and +0.03 [-0.04 to +0.10], intervals consistent with no difference. At 10 m resolution, convolutional neural networks add no benefit over pixel-based XGBoost. Comparisons with four national- and continental-scale products show the importance of non-old-growth labels, and reveal 81% agreement between our predictions and a field-calibrated map. We conclude that buffered spatial validation is vital when transferring old-growth detection models to unseen landscapes, and provide our labels and predictions for future work.

1. Introduction

The study evaluates whether Sentinel features and geospatial foundation model embeddings improve old-growth forest detection beyond topography and access, using high-confidence parcel labels and spatially blocked validation in Romania’s Făgăraș Mountains. It also describes buffered nested validation, calibrated old-growth mapping, and delineation of the area where predictions are expected to apply.

1. Introduction

  • The study evaluates Sentinel-1/2, AlphaEarth and TESSERA v2 features beyond topography and access, and tests whether CNN spatial context improves on pixel-based XGBoost.It also compares predictions with four national- and continental-scale forest maps and provides study-area prediction layers.Source: §1 · p. 2, §2.5 · p. 6

2.1. Study area

  • The 211,893 ha Făgăraș Mountains study area includes 152,366 ha of forest, with broadleaf, mixed and coniferous stands along an elevational vegetation gradient.Its rugged terrain and mixed land use make it a representative and challenging European test case.Source: §2.1 · pp. 2, 4, §1 · p. 2

2.2. Reference label construction

  • Labels comprised 1,016 old-growth and 3,837 non-old-growth parcels; expert-confirmed virgin or quasi-virgin stands informed positive labels, while stands aged 80 years or less defined negatives.Older parcels received no age-defined label and could be classified as either class.Source: §2.2 · p. 4

2.3. Performance metrics

  • Parcel-level precision-recall AUC was the primary metric because it is threshold-free, avoids inflation from the non-old-growth majority, and is sensitive to the old-growth minority class.ROC-AUC, overlap coefficient, precision, recall and F1 complemented it.Source: §2.3 · p. 5

2.4. Evaluation against existing studies

  • Four existing national- and continental-scale forest studies were evaluated against the reference labels using a common parcel-level approach.Their mapped categories differ, so the assessment considered binary and continuous outputs where available.Source: §2.4 · pp. 5–6

2.5. Feature sets and model architectures

  • The five feature sets comprised a coordinate-only control, a six-band topography-and-access baseline, and baseline additions of Sentinel-1/2, AlphaEarth or TESSERA features.Four architectures—single-pixel XGBoost and CNNs with 3 × 3, 5 × 5 or 7 × 7 receptive fields—were tested across the feature sets.Source: Table 2 · p. 5, §2.5 · p. 6

2.6. Spatial autocorrelation

  • Reference labels and predictors showed kilometre-scale spatial dependence: label semivariogram range was 6.6–6.7 km, and Moran’s I was 0.10 at 10 km.Elevation, slope and distance to paved roads had the longest predictor ranges; AlphaEarth and TESSERA median ranges were 9.4 km and 10.5 km, respectively.Source: §2.6 · p. 6, Figure 2 · p. 7

2.7. Spatially blocked nested cross-validation

  • Reference parcels were assigned to six spatially blocked folds, with adjacent parcels merged before assignment and a median cross-fold parcel distance of 22–34 km.Training data near test-fold boundaries were removed to control spatial leakage.Source: §2.1 · p. 3, §2.7 · pp. 6–7
  • Sensitivity analyses found that varying fold composition, training-pixel sampling, model seed, and bootstrap block count did not change the conclusions.The nested design held out each of six spatial folds for testing, with inner-loop hyperparameter tuning; buffering excluded nearby training parcels before refitting models.Source: §2.7 · pp. 7–8

2.8. Mapping

  • The best-performing model produced calibrated 10 m probability and binary rasters, plus calibrated probability and binary classifications for forest parcels.Separate pixel- and parcel-level Platt-scaling models calibrated scores; 15,443 parcels were classified, while 72 parcels without pixel centres were not.Source: §2.8 · p. 8

2.9. Determination of area of applicability

  • The area of applicability (AOA) covered Carpathian pixels whose predictor dissimilarity did not exceed the upper whisker of cross-validated training-pixel dissimilarity.Dissimilarity was measured against the nearest training pixel outside each pixel’s fold, using mean-standardised predictors weighted by XGBoost gain.Source: §2.9 · p. 8

3. Results

Existing forest maps vary substantially in extent and agreement with the study’s reference labels. Earth-observation features improve parcel-level detection, but TESSERA’s unbuffered lead narrows under buffered validation, and CNNs add no parcel-level benefit.

3.1. Evaluation of existing studies

  • Schickhofer and Schwarz’s field-calibrated map ranked highest and Sabatini’s Europe-wide map lowest across PR-AUC, ROC-AUC and F1; mapped extents differed threefold.Schickhofer and Schwarz achieved PR-AUC 0.65, recovering 90% of reference OGF parcels at 64% precision; Sabatini achieved PR-AUC 0.35 against a 0.21 no-skill floor.Source: §3.1 · p. 8, Table 3 · p. 9

3.2. Contribution of Earth observation feature sets

  • Adding conventional EO, AlphaEarth or TESSERA increased PR-AUC by 0.21–0.25 under a 10 km buffer, with 95% CIs of 0.15–0.34.The gains remained when training parcels within 10 km and 20 km of held-out folds were excluded.Source: §3.2 · p. 8, Figure 4 · p. 9
  • TESSERA reached PR-AUC 0.84 [95% CI 0.79–0.88] without buffering; its +0.08 advantage over conventional EO and AlphaEarth was significant, but buffered contrasts were not.At 10 km, the contrasts were +0.04 [−0.01 to 0.11] versus conventional EO and +0.03 [−0.04 to 0.10] versus AlphaEarth.Source: Figure 4 · p. 9, §3.2 · p. 10
  • Distance to unpaved road contributed more gain than any other individual predictor in every feature set, while AlphaEarth and TESSERA embeddings accounted for 0.82 and 0.87 of gain, respectively.The embedding dimensions have no direct physical interpretation.Source: §3.2 · p. 10, Figure 6 · p. 11

3.3. Contribution of spatial context

  • CNNs did not improve parcel-level performance for any feature set; the best baseline-plus-EO CNNs scored 0.05–0.07 PR-AUC below XGBoost.At pixel level, the CNN penalty was less conclusive, with moderate point-estimate improvements in some configurations.Source: Figure 4 · p. 9, §3.3 · p. 10

3.4. Mapping old-growth in the Făgăras, Mountains

  • The baseline-plus-TESSERA XGBoost model classified 55,025 ha [39,124–62,515 ha] as candidate old-growth forest, covering 36.1% [25.7–41.0%] of the forested AOI.Its calibrated pixel raster and parcel vector had in-sample expected calibration errors of 0.5% and 2.0%, respectively.Source: §3.4 · pp. 10–11
  • Predictions agreed with the field-calibrated Schickhofer and Schwarz map on 81% of parcels after excluding parcels with reference labels.Agreement ranged from 67% to 87% across stand-age strata and was higher for parcels predicted as non-OGF than OGF.Source: §3.4 · p. 11

3.5. Area of applicability within the AOI and Carpathians

  • Across the Carpathians, 6.44 Mha (65.0%) of 9.90 Mha of forest fell within the area of applicability; coverage was at least 78% from 700 to 2,000 m.Only 18.3% of forests below 400 m and 32.0% above 2,000 m were within the area of applicability.Source: §3.5 · p. 11

4. Discussion

The discussion finds a spatially robust old-growth signal in satellite data, while showing that buffered validation narrows TESSERA’s apparent advantage. It emphasizes confirmed negatives, spatially diverse labels, and validation matched to the intended transfer setting.

4.1. Overview

  • Existing maps varied in agreement and mapped extent; field-calibrated expert interpretation set the benchmark for the locally trained model.The authors provide raster and vector layers of their reference labels and old-growth predictions.Source: §4.1 · pp. 11–12

4.2. Knowing where old-growth forests are not

  • Presence-only products had low precision and PR-AUC on the reference labels, underscoring the value of confirmed non-old-growth labels for training and assessment.Because old-growth is rare, even a low false-positive rate can sharply reduce precision.Source: §4.2 · p. 12

4.3. EO outperforms simple topographic and access features

  • Optical and radar satellite data contribute predictive skill beyond topography and access, with skill robust to spatial validation and train-test buffering.This supports a pathway to large-scale mapping independent of ALS acquisition.Source: §4.3 · p. 12

4.4. Spatial autocorrelation can inflate performance under blocked validation

  • TESSERA’s unbuffered advantage over conventional EO and AlphaEarth was no longer statistically significant with a 10 km buffer.The authors selected 10 km using residual semivariogram, buffered-control, and coordinate-only diagnostics; they recommend interpreting EO comparisons in light of spatial independence.Source: §4.4 · pp. 12–13

4.5. Spatial context introduced by CNNs adds no benefit at 10 m spatial resolution

  • CNNs did not improve parcel-level performance; at 10 m resolution, spatial context may add no more than the denoising already provided by parcel aggregation.The authors suggest that horizontal structures distinguishing old-growth may be poorly resolved at this pixel size.Source: §4.5 · p. 13

4.6. Limitations

  • Results come from one mountain beech–spruce landscape, and outside-AOI accuracy was not tested or field-verified.The non-probability reference sample also constrains performance estimates and mapped extent, which is not adjusted for omission and commission errors.Source: §4.6 · p. 13

4.7. Implications for Europe-wide mapping

  • Future large-scale mapping needs more diverse labels, low-prevalence-appropriate evaluation metrics, and tests of temporal transfer.The authors identify precision-recall metrics as essential and temporal transfer as a way to monitor forest loss and recovery.Source: §4.7 · p. 13

5. Conclusion

The conclusion presents freely available satellite data as useful for detecting old-growth beyond topographic and access predictors. It stresses buffered validation for transfer to unseen landscapes and identifies reference-label diversity as a remaining constraint.

5. Conclusion

  • EO-derived feature sets were consistent with no difference when a 10 km buffer separated training and test labels, supporting buffered validation for transfer to unseen landscapes.The conclusion also states that spatially blocked validation alone is not sufficient to control spatial autocorrelation.Source: §5 · p. 14
  • 0.84 PR-AUC [95% CI 0.79–0.88] was achieved by the best-performing configuration within the AOI.The conclusion says CNNs showed no advantage over pixel-based XGBoost with 10 m predictors.Source: §5 · p. 14
  • Geographic diversity of reference labels and confirmed negatives remain limiting factors, and the area-of-applicability analysis identifies Carpathian forests where new labels would contribute most.The conclusion also notes that freely available satellite data detect old-growth beyond topography and access predictors.Source: §5 · p. 14

CRediT authorship contribution statement

The statement assigns conceptualisation and writing roles across all four authors, alongside distinct technical and project responsibilities.

CRediT authorship contribution statement

  • All four authors contributed to conceptualisation and manuscript review, while Ratsakatika additionally led methodology, validation, analysis, investigation, data curation, visualization, project administration, and funding acquisition.Zotta contributed data curation, investigation, and supervision; Keshav and Lines contributed methodology and supervision.Source: CRediT authorship contribution statement · p. 14

Data and code availability

The paper makes its reference labels, predictions, and analysis code available, while documenting sources for the underlying and comparison data.

Data and code availability

  • Old-growth reference labels and AOI predictions are available on Zenodo, and the code to download AOI data and reproduce the analysis is available on GitHub.Comparison-study data sources and access conditions are also listed, including vectors that must be obtained from their authors.Source: Data and code availability · p. 14
  • Reference labels and parcel geometries came from datasets compiled by Fundat,ia Conservation Carpathia from forest plans, public records, and ministry applications.Several underlying forest plans have since been superseded.Source: Data and code availability · p. 14

S1. Literature search

The literature search targeted old-growth and related forest concepts alongside convolutional-neural-network and geospatial-foundation-model methods. It was conducted on 18 September 2026, and its relevant-results table was restricted to temperate-forest studies explicitly targeting these concepts.

S1. Literature search

  • 313 results were returned by the Web of Science query conducted on 18 September 2026.Source: S1. Literature search · p. 20
  • The query combined old-growth, primary, naturalness and related forest terms with convolutional-neural-network, deep-learning and geospatial-embedding terms.The search also included named methods and models such as AlphaEarth and TESSERA.Source: S1. Literature search · p. 20
  • Table S1 restricts relevant results to temperate-forest studies where old-growth, primary or similar concepts are the explicit prediction target.Source: Table S1 · p. 20

S2.1. Reference label construction: additional details

The reference parcel geometry was cleaned, enriched with forest-management attributes and filtered to support classification. The supplement reports exclusions from the label analysis and final counts for old-growth and non-old-growth parcels.

S2.1. Reference label construction: additional details

  • The base geometry comprised 16,199 forest parcel boundaries and was the only accessible contiguous parcel source across most of the study area, despite revisions since compilation around 2018.Source: S2.1. Reference label construction:… · p. 20
  • The parcel layer was de-overlapped and filtered, leaving 15,515 clean parcels and 15,443 classified parcels after excluding 72 without pixel centres.The procedure also removed 506 parcels outside the area of interest and 96 nonproductive units.Source: S2.1. Reference label construction:… · p. 21
  • Management-plan attributes were matched to parcels at 90% geometry overlap, and final parcel age used the dated record closest to 2020, preferring the earlier record in a tie.Attributes included species composition, average age and management-plan year.Source: S2.1. Reference label construction:… · p. 21
  • The label analysis excluded 66 parcels without CORINE classifications and six without predictor-mask pixels, leaving 1,006 old-growth and 3,775 non-old-growth parcels.Source: Figure S1 · p. 21

S2.2. Existing study OGF fraction sensitivity

The supplement describes how the existing-study parcel-fraction threshold was selected and how spatial-dependence ranges were interpreted. It identifies the 0.5 majority rule and criteria for reporting unresolved ranges and selecting a best-fit model.

S2.2. Existing study OGF fraction sensitivity

  • The 0.5 majority rule was the selected threshold for existing studies, and Fig. S2 plots each product’s F1 across parcel OGF-fraction thresholds.Source: Figure S2 · p. 21

S2.3. Label and predictor autocorrelation analysis

  • A spatial-dependence range was unresolved if it reached 90% of the 40 km maximum lag or fell below the 1 km first lag class; the best model was the resolved fit with lowest RMSE.Moran’s I values were significant at p < 0.05 except those marked with a dagger.Source: S2.3. Label and predictor… · p. 22, Table S2 · p. 22

S2.4. Fold and bootstrap block balance

The supplement reports balance of total area across spatial folds and bootstrap blocks. It notes that old-growth-area balance was not optimized for the bootstrap blocks.

S2.4. Fold and bootstrap block balance

  • Fig. S3 presents the balance of total and old-growth area across folds.Source: Figure S3 · p. 23
  • Fig. S4 presents total-area balance across bootstrap blocks; the optimization objective used area only, not old-growth balance.Source: Figure S4 · p. 23

S2.5. Sensitivity analysis: block count, hyperparameter search, pixel sampling method

The sensitivity analyses examine how block count, hyperparameter search behavior, and pixel-sampling choices relate to model performance. Their tables report pooled and per-fold PR-AUC, with confidence intervals for block-count sensitivity.

Block count

Hyperparameter search

  • Table S4 summarizes the spread of inner-fold parcel PR-AUC across hyperparameter trials and search flatness near the optimum.Values are summarized as median [min, max] across configuration-by-outer-fold combinations.Source: Table S4 · p. 23

Pixel sampling method

  • Table S5 compares pooled and per-fold out-of-fold performance under alternative pixel-sampling modes at parcel and pixel levels.It covers the baseline + TESSERA XGBoost model; F1 uses the pooled out-of-fold predictions’ F1-maximizing threshold.Source: Table S5 · p. 24

S3. Results

The supplementary results report parcel- and pixel-level performance across feature sets, architectures, spatial buffers, and decision thresholds. They also describe paired contrasts, robustness checks, and comparisons with existing forest products.

S3. Results

  • Pixel-level out-of-fold performance is tabulated for each feature set and architecture at a 0 km training-exclusion buffer.The table distinguishes the pooled outer F1-maximizing threshold from inner-fold thresholds.Source: Table S6 · p. 24
  • Additional supplementary analyses show overlap distributions for TESSERA predictions against continuous primary-forest and structural-complexity surfaces, and bootstrap old-growth area estimates.The area bootstrap uses 1,000 replicates and reports a point estimate, bootstrap mean, 95% CI, and existing study estimates.Source: Figure S5 · p. 24, S3. Results · p. 24, Figure S8 · p. 33
  • Pooled out-of-fold parcel-level performance is reported by feature set, architecture, and training-exclusion buffer, at threshold-free and two decision-threshold settings.The performance measures include PR-AUC, ROC-AUC, F1, precision, and recall.Source: S3. Results · pp. 25–26, Table S7 · p. 25
  • Paired contrasts compare feature sets against baseline and coordinate-only predictors, and compare Earth-observation feature sets, across training-exclusion buffers.The section also reports paired CNN receptive-field contrasts against XGBoost by feature set and buffer.Source: S3. Results · pp. 26–27, Table S8 · p. 26, Table S9 · p. 26
  • Robustness checks cover per-fold PR-AUC, ROC-AUC, and F1, leave-one-fold-out PR-AUC and paired contrasts, seed variability, and residual spatial dependence.The spatial buffering figures also describe relative performance differences, training-data loss, OGF prevalence, and PR-AUC for spatially gapped and parcel-matched controls.Source: Table S11 · p. 28, Table S12 · p. 29, Table S13 · p. 30
  • The supplementary material includes mapped old-growth area estimates, stratified ROC-AUC comparisons, and agreement assessments against other forest products.It includes a cross-study parcel-level agreement figure and a table comparing the deployed model with Schickhofer and Schwarz by labelling stratum.Source: Table S19 · p. 33, Table S20 · p. 33, Table S21 · p. 33

Related papers

Loading 2609.28194v1…