The conservation value of forests can be predicted at the scale of 1 hectare

The study addresses the urgent need to map and prioritize forest areas with high naturalness and biodiversity conservation value (High Conservation Value Forests, HCVF) to support functional habitat networks and green infrastructure. Intensive forestry has transformed much of Sweden’s naturally dynamic forests into wood production systems, causing loss of structural complexity and habitat continuity. Existing national HCVF data, compiled from decades of field surveys without a predefined sampling scheme, are incomplete, outdated, and spatially biased. Leveraging newly available wall-to-wall spatial datasets and advances in machine learning, the authors aim to produce validated, high-resolution (1 ha) predictions of the relative likelihood of HCVF occurrence across Sweden. The research questions are whether open spatial data and Random Forest models can reliably predict gradients of forest naturalness and locate HCVF, and how such maps can inform conservation targets, spatial planning, and forest landscape restoration.

The paper situates its work among efforts to identify forests with high naturalness using field inventories (e.g., woodland key habitats), historical and contemporary databases, and remote sensing with machine learning. Global and continental products (e.g., Intact Forest Landscapes, European primary forests) are often too coarse or trained on spatially clustered data, limiting local planning utility. Recent regional studies (e.g., Munteanu et al. in the Carpathians using Maxent; Ørka et al. in Norway using boosted regression for LiDAR-based naturalness) share conceptual similarities but differ: prior work often excludes historically disturbed forests, provides binary outputs, uses limited predictors, or omits multi-scale drivers and direct human impact proxies. The authors argue that regionalized modelling, multi-scale variables, and integration of both forest structure and anthropogenic proxies can yield actionable, landscape-scale predictions suitable for national conservation planning.

Study area and stratification: Sweden was divided into four biogeographic/administrative regions (North boreal, South boreal, Hemiboreal, Nemoral). The modelling unit was 1 ha (100×100 m) pixels masked to forest-dominated pixels (forest cover > 0.5). Predictions covered ~21.85 million 1 ha forest pixels across the four regions (~78% of Sweden’s forest area). Data sources: Predictor data were drawn from publicly available datasets: National Land Cover Data (NMD; 10 m; forest types, productivity, logged/temporary non-forest, broadleaf share, forest cover), LiDAR-derived forest structure (tree height, height variance, understory, gaps), DEM (50 m elevation, slope), Global Forest Change (GFC; 30 m loss 2000–2020 and gain 2000–2012), harmonized night-time lights (1 km), and population (1 km). Distance layers to roads, built-up areas, and water were computed. All layers were resampled/aggregated to 1 ha using appropriate statistics (average, variance, proportion), preserving thematic information. Two Shannon diversity indices (forest types; natural land cover) were computed. Tree height was regionalized (HEIGHTC) as deviation from mean height per forest type within a 10×10 km window to account for climatic/productivity gradients. Multi-scale features: Selected variables were summarized using moving window aggregation at 0.3, 0.5, 1.1, 5.1, and 10.1 km window sizes (e.g., HANSEN003/005/011/051/101; FOPEN003/011; DEM011; SLOPE011; BROADLEAF005; SHAFOR003). The initial predictor set (including multi-scale variants) comprised 128 variables; 31 base predictors are listed, with multi-scale expansions where applicable. Response and sampling: Presence-only HCVF data came from the national Swedish HCVF database (641,095 polygons; ~3.44 Mha; updated to 2019–2020 for mountain region). Polygons were rasterized at 1 ha, selecting pixels with HCVF proportion ≥ 0.5, excluding clusters <10 ha. Presence samples were selected with ≥5 km spacing. Pseudo-absences were sampled from forest-masked areas buffered ≥1 km away from any HCVF polygon, excluding clusters <10 ha, and spaced ≥5 km, yielding a moderately imbalanced dataset. Variable selection: To reduce collinearity and enhance interpretability, univariate Random Forests were fitted for all predictors; variables were ranked by mean ROC AUC and Pearson correlation from 10-fold spatial cross-validation (SCV). Pairs with |r|>0.7 were filtered, retaining the better-scoring variable. Final models used 49–53 predictors per region (North boreal: 49; South boreal: 50; Hemiboreal: 53; Nemoral: 48). Modeling: Regional Random Forest (RF) classifiers were trained independently per region using scikit-learn (n_estimators=500) with a balanced RF implementation (downsampling within trees) to handle class imbalance. Hyperparameter tuning via Bayesian optimization confirmed default settings were robust (<0.5% ROC AUC gain). Performance was assessed with 10-fold SCV using 20×20 km spatial blocks. Threshold-independent metrics emphasized continuous outputs: ROC AUC, PR AUC, Brier score, Pearson’s r; also Accuracy, TSS, MCC at a 0.5 threshold. External validation: Two independent datasets validated predictions: (1) Sveaskog stand-level data (n=57,548 stands) with management objective categories (NF, NF_NM, NM, PE, PG) and a binary naturalness label; per-stand mean predicted likelihood computed over ≥10 pixels. (2) Swedish National Forest Inventory plots (n=13,775; 2015–2019) with naturalness classes (plantation, normal, natural) and Natura 2000 habitat indicator; mean predicted likelihood extracted from the plot pixel and four nearest neighbors. Group differences were tested with Tukey’s posthoc tests. Comparative models: Robustness checks compared the selected regional model to global (unstratified), full (all predictors), one-scale (no multi-scale), baseline (four predictors), and Logistic Regression alternatives. Pixel-wise correlations and external validation assessed spatial and predictive differences. Mapping: Final regional RF models were retrained on all training data and applied wall-to-wall to produce a national 1 ha map of relative likelihood of HCVF occurrence, with partial dependence plots and impurity-based variable importance summarizing drivers.

- Predictive performance (10-fold spatial CV): Across regions, the RF models achieved ROC AUC 0.89–0.90, PR AUC 0.84–0.89, Pearson r 0.66–0.69, Brier score 0.14–0.15, MCC 0.61–0.62, Accuracy ~0.81, TSS 0.57–0.61, indicating strong predictive capability for continuous relative likelihoods of HCVF. - Important predictors: Consistently influential variables included LiDAR-based forest structural metrics (HEIGHT, HEIGHTC, HEIGHTV; multi-scale height summaries), management intensity proxies (GFC loss/gain: HANSEN003/005/011; NMD logged/temporary non-forest: FOPEN003/011), and topography (DEM011, SLOPE011). In southern regions, BROADLEAF proportion contributed positively. Partial dependence showed monotonic increases of likelihood with taller-than-average and structurally diverse stands, steeper slopes, and higher elevations; monotonic decreases with greater management intensity, road proximity/density, and built-up influence. - Spatial coverage: Predictions generated for ~21.85 million 1 ha forest pixels, representing ~78% of Sweden’s forest area, enabling national wall-to-wall ranking of forest naturalness. - External validation (Sveaskog stands, n=57,548): Mean predicted likelihoods followed the ordinal management gradient. NF (non-intervention) and NF_NM (conservation, not yet specified) were highest (region means: NF 0.64/0.67/0.59/0.72; NF_NM 0.72/0.70/0.67/NA), followed by NM (0.57/0.53/0.58/0.63), then PE (0.42/0.40/0.43/0.59), and lowest PG (0.29/0.28/0.32/0.37). Conservation-oriented vs production-oriented categories differed highly significantly (p<0.001) in all regions except some overlaps in Nemoral. Stands labeled natural had significantly higher means (0.65/0.63/0.58/0.65) than non-natural (0.30/0.30/0.33/0.39). - External validation (NFI plots, n=13,775): Natural plots had higher means (North/South/Hemiboreal: 0.84/0.76/0.63) than normal (0.36/0.37/0.38) and plantations (0.20/0.24/0.27); Nemoral had too few natural plots for comparison. Plots meeting Natura 2000 habitat criteria had markedly higher means (0.83/0.83/0.72/0.69) than those without (0.35/0.36/0.37/0.41), with high statistical significance. - Model comparisons: Regionalized RF matched or outperformed alternatives; global model performed similarly overall but differed spatially in southern regions. One-scale models were slightly weaker; baseline models weakest. RF outperformed Logistic Regression by ~2–4% in ROC/PR AUC in most regions; Hemiboreal was similar across methods. - Ecological patterns: HCVF likelihood concentrated in complex terrain, higher elevations, taller/heterogeneous stands, and areas with low recent management intensity and accessibility; multi-scale predictors captured both forest-dominated and isolated HCVF patches in non-forest matrices.

The results demonstrate that integrating open spatial datasets with Random Forest modelling can produce reliable, fine-grained predictions of forest naturalness that align with independent, field-verified indicators of conservation value at stand and plot scales. The continuous likelihood surface addresses the need to prioritize areas along a gradient—from intact/high-naturalness forests for strict protection to intermediate areas suitable for restoration to enhance connectivity and functionality of green infrastructure. Compared to global or local niche mappings, the regionalized, multi-scale approach captures spatial heterogeneity and complex, non-linear relationships, yielding actionable outputs for national to landscape-scale planning. The maps can inform zoning (e.g., TRIAD), identify low-conflict areas for production-focused management, and target restoration to expand or connect existing hotspots. The study underscores the advantage of combining structural LiDAR metrics with human impact proxies and multi-scale context, and supports regional stratification to address spatial non-stationarity.

The study delivers a validated, wall-to-wall 1 ha map of the relative likelihood of High Conservation Value Forests across Sweden using open data and machine learning. By ranking forests along a continuous naturalness gradient, the outputs fill a critical gap for evidence-based conservation target assessment, spatial planning, and forest landscape restoration. The approach is robust across eco-regions and transferable in principle where comparable data exist. Future work should focus on expanding and improving training/reference datasets (e.g., national inventories, primary forest databases, species records), integrating additional structural proxies where LiDAR is limited (e.g., satellite-derived canopy height), advancing spatially explicit learning methods to handle heterogeneity, and applying/validating the framework in other countries and forest biomes. Field verification remains essential before operational decisions.

- Training data limitations: The HCVF database is presence-only, compiled over decades without probabilistic sampling; it mixes protected, set-aside, and unprotected areas and may include variability in actual conservation value. Failed field verifications of candidate HCVFs were not recorded. Consequently, models predict relative likelihoods, not true probabilities, and some noise is expected. - Sampling bias and class imbalance: Historical mapping and socio-economic factors likely introduced spatial bias; the study mitigated this using spacing rules, buffers, and balanced RF but residual effects may persist. - Predictor limitations: GFC loss/gain captures forest cover continuity but does not distinguish causes (natural vs harvest). Availability and resolution of LiDAR-derived structure vary; in regions lacking LiDAR, substitutes may reduce fidelity. - Transferability and data availability: Application to other regions depends on access to comparable wall-to-wall predictors and reliable HCVF training data; LiDAR scarcity and limited reference datasets constrain scalability. - Regional data sparsity: Some validation categories (e.g., Nemoral “natural” plots) had very few observations, limiting inference locally. - Interpretability: RF outputs represent fractions of tree votes (probability-like) and are not calibrated probabilities; multi-collinearity was reduced but complex interactions limit mechanistic interpretation.