A 30 m forest dominant height dataset for China in 2020 Authors: Yuling Chen, Guangcai Xu, Haitao Yang, Qinghua Guo Published: 2026 ABSTRACT AbstractForest dominant height is a fundamental structural attribute that reflects site conditions and forest growth potential. Here we present a nationwide 30 m resolution forest dominant height dataset for China (FDH-30C). The dataset is calibrated using 1,117 km² of high-density unmanned aerial vehicle (UAV) light detection and ranging (LiDAR) data distributed across all eight major vegetation divisions across China as reference data, representing diverse stand ages, structures, and species compositions. To produce spatially continuous estimates, the model used in this dataset integrates 30 geospatial predictors derived from multi-source remote sensing products, including climatic, edaphic, topographic, vegetation, and Synthetic Aperture Radar (SAR)-based variables. A two-stage hybrid modeling framework combines the UAV LiDAR reference data with these predictors to generate spatially coherent estimates while preserving local accuracy and reducing ecozone boundary effects. The resulting map provides a consistent national baseline for applications such as site-index mapping, growth-and-yield parameterization, biomass and carbon estimation, vertical structure analysis, and the evaluation of spaceborne LiDAR missions. Background & Summary Accurate quantification of forest structure is essential for understanding resource availability and guiding management decisions1. Among structural parameters, tree height and diameter at breast height (DBH) are the two most widely used indicators of stand productivity and biomass2,3. Yet, compared with DBH, the accurate acquisition of tree height remains technically challenging due to canopy occlusion, terrain complexity, and limitations of ground-based observations4,5. Within stand-level or plot-level assessments, forest height can be summarized as either mean height or dominant height (hd)6. Mean height provides a general indication of stand development, whereas hd reflects the growth conditions of the most vigorous trees and serves as a robust proxy for site productivity7,8. Because dominant trees are less affected by understory competition and management interventions, their height responds primarily to site factors, making forest dominant height a key metric for evaluating forest growth potential, site quality, and long-term yield capacity8–12. Despite its central role in forest mensuration and modeling, hd has rarely been characterized with spatial consistency across large regions. Existing definitions—whether based on the tallest trees or those with the largest diameters—further complicate its estimation and comparison among studies13. However, identifying dominant trees by diameter is less practical for large-scale applications, as accurate DBH measurements are limited to field inventories14. Consequently, reliable, spatially explicit information on hd remains scarce, restricting its broader integration into forest resource assessments and ecosystem modeling. Light detection and ranging (LiDAR) has transformed forest structural assessment by directly capturing three-dimensional canopy and stem information through active laser scanning15–20. Recent advances in close-range platforms—particularly unmanned aerial vehicle (UAV) laser scanning—have greatly improved spatial resolution, data accessibility, and cost efficiency, enabling reliable measurement of individual trees across diverse forest types20–22. LiDAR-based tree height estimation can generally follow two approaches: an area-based method17,23, which generates a canopy height model to calculate tree or canopy height by aggregating LiDAR-derived metrics that characterize the vertical distribution of returns within predefined spatial units, such as plots or grid cells, and a tree-based method19,22,24, which measures height at the individual tree level. While area-based metrics are computationally convenient, they often mix vegetation and non-vegetation elements, deviating from the forestry definition of tree height23. In contrast, tree-based estimation aligns more closely with ecological reality by identifying individual trees, quantifying their heights, and aggregating these measurements to represent stand structure22. Such an approach, supported by the growing availability of UAV LiDAR data, allows for accurate and consistent assessment of forest vertical structure over extensive areas. With the accumulation of large-scale UAV LiDAR observations across China, this study establishes a solid foundation for mapping hd in a spatially continuous manner. However, the vast geographic extent and environmental heterogeneity of China’s forests introduce a critical challenge—how to systematically integrate these massive, site-specific measurements to generate accurate, nationally consistent hd products for ecological assessment and forest management. Producing a wall-to-wall hd product for China requires addressing two fundamental challenges. First, forest structure varies substantially across ecozones, raising the question of how to balance regional accuracy with national consistency. While zone-specific models generally improve local performance, they can introduce discontinuities at ecological boundaries25. Rather than seeking to identify a universally superior modeling approach, this study adopts both strategies as complementary pathways for dataset construction, investigating unified frameworks where ecozone variability is treated as a random effect—such as in mixed-effects formulations—to evaluate whether a single model can retain coherence and adaptability26. The second constraint arises from the spatial coverage of close-range LiDAR sampling, which, despite their accuracy at the plot level, remain insufficient to support direct large-scale surface construction. Extending these measurements to unsampled regions is achieved through the integration of multi-source remote sensing predictors within machine learning frameworks, enabling the construction of a spatially continuous hd product17,27–29. These extensive, tree-based measurements form the basis for evaluating strategies to generate continuous, nationally consistent hd product. To address the absence of spatially consistent stand-height information, this study utilizes more than 1,000 km² of UAV LiDAR data collected across eight major vegetation divisions in China. Forest dominant height is defined as the mean height of the three tallest trees within each 30 m grid, ensuring alignment with operational mensuration practices. Based on these tree-level measurements, we produced national-scale hd maps using a unified machine learning–based mixed-effects modeling framework with ecozone-level random effects, which explicitly captures hierarchical dependencies among vegetation divisions and improves large-scale spatial continuity. By documenting both the construction and validation of these products, this data paper aims to provide an openly accessible, standardized dataset of hd for China. To our knowledge, this is the first dataset to characterize hd at a national scale using tree-based UAV LiDAR observations. This map is intended to support forest inventory, structural monitoring, carbon assessment, and habitat evaluation, providing a foundational dataset for ecological research and management. Methods Figure 1 illustrates the overall workflow, which consisted of four primary steps: first, processing close-range LiDAR data to segment individual trees and compute forest dominant height; second, constructing a comprehensive feature set from multi-source remote sensing and ancillary data; third, performing machine learning–based mixed-effects modeling; and fourth, mapping forest stand mean height wall-to-wall across China and conducting accuracy assessment. Workflow for producing wall-to-wall maps of forest dominant height at 30 m resolution across China. UAV LiDAR data collection and processing The UAV LiDAR data was collected over an eight-year period ending in 2023 and cover 1,117.76 km² across all major vegetation divisions of China, with the exception of the Qinghai-Tibet Plateau alpine vegetation divisions (Table S1). Approximately 400 TB of point-cloud data were acquired using a uniform UAV LiDAR acquisition protocol (Table 1), forming the core reference dataset for deriving forest dominant height samples. Table 1Percentage of forest area covered by UAV LiDAR data in each vegetation divisions across China.Vegetation divisionTotal area (km2)Percentage of forest area covered by UAV LiDAR data ( × 10 −5)Temperate desert0.180.75Temperate steppe76.0498.16Subtropical evergreen broadleaf forest755.9154.28Tropical monsoon forest-rainforest65.4239.64Warm temperate deciduous broadleaf forest163.9774.92Temperate needleleaf-broadleaf mixed forest53.6119.46Cold temperate needleleaf forest2.631.16Note: The forest area across China is derived from Cheng, K. et al.33. Percentage of forest area covered by UAV LiDAR data in each vegetation divisions across China. Note: The forest area across China is derived from Cheng, K. et al.33. All point clouds were processed through a standardized workflow implemented in LiDAR360 (Version 6.0, www.lidar360.com), including resampling, denoising, ground point classification, and normalization. Individual trees were segmented using the Li et al.30 algorithm embedded in LiDAR360. To account for substantial variation in forest structure across China, segmentation parameters were optimized at the file level. For each LiDAR file, a representative local subset (typically 5 × 5 plots) was selected for parameter tuning. The key parameter, Sigma, which controls Gaussian smoothing, was tested from 0.2 to 2.0 at 0.2 intervals. Lower values were used to reduce under-segmentation, whereas higher values were used to reduce over-segmentation. The optimal Sigma value was determined by visual inspection of the subset results and then applied to the full file. For some early LiDAR datasets, where individual trees could not be reliably identified visually, Sigma selection was additionally informed by stand density or related structural information reported for the corresponding region. This procedure allowed the segmentation settings to be adapted to differences in vegetation structure while maintaining a consistent workflow for extracting the three tallest trees within each 30 × 30 m plot. Following the tree-level segmentation, the LiDAR data covering irregular large regions were divided into 30 × 30 m grid cells representing forest plots. Only grid cells fully contained within the stand boundaries were retained as valid plot samples. Before plot generation, additional filtering was applied to ensure sample consistency and data quality. LiDAR data without valid geographic coordinates and non-representative urban or fragmented forest patches were excluded. For large LiDAR forest files, only internally complete 1 km × 1 km blocks were retained before subdivision into 30 m × 30 m plots, with incomplete edge and corner areas removed. Small LiDAR files that could not support a complete 30 m × 30 m plot were also excluded. This approach minimized edge effects and guaranteed that each plot represented complete and continuous LiDAR coverage. In total, 610,342 valid sample plots were generated. Forest dominant height definition Forest dominant height has been defined in several ways6,31,32, including: (1) the average height of the top 100 trees per hectare; (2) regression-based estimates corresponding to the quadratic mean diameter of the 20% thickest trees; and (3) the mean height of the three to six tallest trees in a plot. Consistent with conventional Chinese forest survey methods32, we applied the third definition in Eq. (1), calculating dominant height as the mean height of the top three trees in each 30 × 30 m plot. where hd is forest dominant height (m), hi is tree height of i-th tree (normally hi ≥ 2.0 m is adopted). Feature data processing To support nationwide modeling of hd, 30 predictors were extracted from multi-source remote sensing, classified into five categories: climatic, edaphic, topographic, Landsat-derived optical, and SAR predictors (Table 2). Climate predictors were sourced from the WorldClim 2.1 database (www.worldclim.org), which provides 19 bioclimatic predictors at 30-arc-second resolution. Edaphic attributes, including soil type, soil depth, and texture, were extracted from the Harmonized World Soil Database (HWSD, Version 1.2), with the China component derived from the 1:1,000,000 Soil Map of China provided by the Institute of Soil Science, Chinese Academy of Sciences. Topographic metrics, including elevation, slope, and aspect, were computed from the Shuttle Radar Topography Mission (SRTM) Version 3 (SRTM Plus) product. Vegetation conditions were captured through the annual maximum Normalized Difference Vegetation Index NDVI composite, extracted from Landsat data using Google Earth Engine (GEE). SAR variables, including VV and VH backscatter coefficients, were generated from Sentinel-1 observations. Additionally, a national forest type map distinguishing planted and natural forests33 and a forest age map for China in 202034,35 were incorporated. All layers were harmonized to 30-m resolution using nearest-neighbor resampling to maintain categorical integrity. Table 2Features for mapping forest dominant height in China.Data typeData sourceResolutionTime periodPredictorsForest maskPlanted and natural forest map3330 m2020Planted and natural forestForest ageForest age map3530 m2020Forest ageSAR dataSentinel-130 m2015–2020VV, VHLandsat-derived optical dataLandsat30 m2015–2020NDVI_MAXClimate dataWorldClim V2.130 arc-second1970–2000BIO1-BIO19Soil dataHWSD V1.230 arc-second2007–2009SU_SYM90, REF_DEPTH, T_TEXTURETopographic dataSRTM DEM30 m2000Aspect, Elevation, Slope Features for mapping forest dominant height in China. Mapping wall-to-wall forest dominant height of China The modeling data are hierarchically structured, with eight vegetation divisions, each containing its corresponding plots. Within-division plots are not statistically independent, and substantial heterogeneity exists among vegetation divisions. To accommodate this hierarchical and heterogeneous structure, we employed a two-stage hybrid modeling framework that integrates machine learning–based nonlinear prediction with a linear mixed-effects model to explicitly account for eco-regional heterogeneity. This method not only improves modeling accuracy but also delivers detailed insights into complex hierarchical data structures. The machine learning-mixed effects (MLME) model in Eq. (2) is composed of two distinct parts: a fixed-effects component and a random-effects component. where MLXi represents the predicted value from the machine learning model for each plot t within vegetation division i. β0 and β1 denote the intercept and the coefficient for the predicted value of MLXi. Zi is the matrix of first-order derivatives for the random effects. bi is the random-effect vector specific to the vegetation division i. Using the predicted values from ML model, the influence of covariates is integrated into the mixed-effects model, accounting for the specific hd within different vegetation divisions. The original dataset was randomly split into two groups: two-thirds were added to the training subset and the remainder to the validation subset. The MLME model was developed through a five-step procedure: Algorithm selection: PyCaret (https://pycaret.org) was used to identify the top 3 machine learning candidate algorithms for hd estimation based on the full dataset. Algorithm selection: PyCaret (https://pycaret.org) was used to identify the top 3 machine learning candidate algorithms for hd estimation based on the full dataset. Hyperparameter optimization: Selected algorithms were tuned using Bayesian optimization via Optuna36 on the training subset (2/3 of data). Hyperparameter optimization: Selected algorithms were tuned using Bayesian optimization via Optuna36 on the training subset (2/3 of data). Model evaluation: Performance of the tuned models was assessed on the validation subset (1/3 of data) using coefficient of determination (R², Eq. (3)), mean square error (MSE, Eq. (4)), root mean square error (RMSE, Eq. (5)), and mean absolute error (MAE, Eq. (6)) to identify the optimal ML algorithm. Model evaluation: Performance of the tuned models was assessed on the validation subset (1/3 of data) using coefficient of determination (R², Eq. (3)), mean square error (MSE, Eq. (4)), root mean square error (RMSE, Eq. (5)), and mean absolute error (MAE, Eq. (6)) to identify the optimal ML algorithm. Vegetation division-specific prediction: The best-performing model was used to generate hd estimates for each vegetation division by optimal ML algorithm using Eq. (2). Maximum likelihood estimation of mixed-effects parameters was performed with the lmer function in the R lme4 package (version 4.3.0). Vegetation division-specific prediction: The best-performing model was used to generate hd estimates for each vegetation division by optimal ML algorithm using Eq. (2). Maximum likelihood estimation of mixed-effects parameters was performed with the lmer function in the R lme4 package (version 4.3.0). Specification of random effects and model assessment: Random effects were specified as (1 | zone) for intercepts and (x | zone) for slopes at the vegetation zone level. Model performance was further evaluated using Akaike information criterion (AIC, Eq. (7)) and Bayesian information criterion (BIC, Eq. (8)) to assess the accuracy and generalizability of MLME parameter estimates. Specification of random effects and model assessment: Random effects were specified as (1 | zone) for intercepts and (x | zone) for slopes at the vegetation zone level. Model performance was further evaluated using Akaike information criterion (AIC, Eq. (7)) and Bayesian information criterion (BIC, Eq. (8)) to assess the accuracy and generalizability of MLME parameter estimates. where yi represents the i-th forest dominant height observed values; yi∧ is the i-th predicted forest dominant height value; n is the number of plots, yi¯ is the mean value for the forest dominant height observed values, L is the value of log-likelihood, K is the number of parameters in the model. Uncertainty analysis The total uncertainty associated with pixel-level hd estimates was partitioned into three independent components, each expressed as a relative percentage following Saatchi et al.37. These components capture errors arising from tree segmentation, forest age estimation, and model prediction. The process of uncertainty analysis was as follows: Measurement uncertainty (εmeasurement): Measurement uncertainty reflects the uncertainty introduced by individual tree segmentation from UAV LiDAR data. It was estimated as the complement of the F1 score (1 − F1) from the segmentation assessment. The mean F1 score was 0.9030, corresponding to a measurement uncertainty of approximately 10%. Measurement uncertainty (εmeasurement): Measurement uncertainty reflects the uncertainty introduced by individual tree segmentation from UAV LiDAR data. It was estimated as the complement of the F1 score (1 − F1) from the segmentation assessment. The mean F1 score was 0.9030, corresponding to a measurement uncertainty of approximately 10%. Product uncertainty (εproduct): Product uncertainty focused on the uncertainty associated with the key predictor forest age estimation. The relative uncertainty for each pixel is calculated using Eq. (9), and then the maximum of mean relative error is calculated using Eq. (10).9εi=RMSEyˆi10ε®max=RMSEy®×100%where RMSE is computed from validation data, yˆi represents the predicted age of for pixel i, y® is the mean forest age. Product uncertainty (εproduct): Product uncertainty focused on the uncertainty associated with the key predictor forest age estimation. The relative uncertainty for each pixel is calculated using Eq. (9), and then the maximum of mean relative error is calculated using Eq. (10).9εi=RMSEyˆi10ε®max=RMSEy®×100%where RMSE is computed from validation data, yˆi represents the predicted age of for pixel i, y® is the mean forest age. Prediction uncertainty (εprediction): Prediction uncertainty both sampling effects and model-based prediction errors. For each pixel, relative uncertainty of forest dominant height was calculated as shown in Eq. (11).11εprediction=RMSEhhˆiwhere RMSEh is the RMSE of the forest dominant height validation set, hˆi is the predicted dominant height value for pixel i. Prediction uncertainty (εprediction): Prediction uncertainty both sampling effects and model-based prediction errors. For each pixel, relative uncertainty of forest dominant height was calculated as shown in Eq. (11).11εprediction=RMSEhhˆiwhere RMSEh is the RMSE of the forest dominant height validation set, hˆi is the predicted dominant height value for pixel i. Propagation of uncertainties to pixel-level height estimates: Assuming the three uncertainty components are independent and random, the total relative uncertainty for pixel i was obtained by error propagation in Eq. (12).12εhi=εmeasurement2+εproduct2+εprediction2 Propagation of uncertainties to pixel-level height estimates: Assuming the three uncertainty components are independent and random, the total relative uncertainty for pixel i was obtained by error propagation in Eq. (12).12εhi=εmeasurement2+εproduct2+εprediction2 National-scale uncertainty in hd was derived by propagating pixel-level uncertainties using Eq. (13). National-scale uncertainty in hd was derived by propagating pixel-level uncertainties using Eq. (13). where N is the number of forest pixels within the national boundary, and hˆi and εhi are the forest dominant height and its relative uncertainty at pixel i, respectively. Data Records The dominant height of China’s forests dataset (FDH-30C) generated is publicly available in Zenodo (https://doi.org/10.5281/zenodo.18286069)38. The repository contains the main raster product, FDH-30C.tif, which stores the 2020 forest dominant height estimates for China at 30 m spatial resolution, together with associated metadata documentation. The raster is provided in GeoTIFF format and can be opened in standard GIS and remote sensing software. The released raster contains one primary variable, forest dominant height, expressed in meters. Each pixel represents a 30 × 30 m ground area. Pixel values indicate the predicted dominant height of forest stands, while non-forest areas and unavailable pixels are encoded as NoData. The dataset provides national coverage for China and represents forest conditions for 2020. The product was developed using LiDAR-derived dominant height observations collected between 2015 and 2023 and a set of environmental predictor variables representing climatic, soil, topographic, vegetation, and SAR conditions. These predictor variables were used for model construction and mapping, but the released dataset itself contains the final forest dominant height layer rather than the full predictor stack. The accompanying metadata file describes the dataset contents, spatial coverage, temporal reference, spatial resolution, variable definition, file format, and data usage information, thereby facilitating data reuse in forest monitoring, ecosystem assessment, and related large-scale ecological applications. Technical Validation Prediction accuracy of MLME model To ensure the reliability and accuracy of the forest dominant height dataset, a multi-stage technical validation was conducted based on ML and mixed-effects modeling. A total of 18 ML algorithms were initially tested using the open-source AutoML framework PyCaret to identify suitable models for forest dominant height estimation. As shown in Table S2 Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LGBM), and eXtreme Gradient Boosting (XGBoost) exhibited superior performance. Extra Trees and Random Forest models were excluded due to overfitting and excessive training time, respectively. Bayesian optimization was then applied for hyperparameter tuning (Table 3), and LGBM was identified as the best-performing algorithm (R² = 0.7901, RMSE = 3.43 m, MAE = 2.48 m; Table 4). Table 3The optimal hyperparameter parameter values based different ML algorithms for hd model.AlgorithmPython PackageHyperparametersXGBoostXGBRegressor(‘min_child_weight’: 1, ‘gamma’: 0.06728229770511852, ‘alpha’: 0.21304909691832852, ‘lambda’: 0.00013819888953716042, ‘colsample_bytree’: 0.4374333788046241, ‘subsample’: 0.6068106161836021, ‘random_state’: 2024)LGBMlightgbm(‘min_child_samples’: 5, ‘max_depth’: 19, ‘learning_rate’: 0.05, ‘colsample_bytree’: 0.4886694204822875, ‘n_estimators’: 7661, ‘cat_smooth’: 33, ‘cat_l2’: 13, ‘min_data_per_group’: 105, ‘cat_feature’: 50, ‘random_state’: 2024)CatBoostcatboost(‘max_bin’: 384, ‘min_data_in_leaf’: 12, ‘l2_leaf_reg’: 0.052482360271174594, ‘subsample’: 0.799227991371045, ‘random_state’: 2024 ‘max_bin’: 384, ‘min_data_in_leaf’: 12, ‘l2_leaf_reg’: 0.052482350271174594, ‘subsample’: 0.799227991371045, ‘random_state’: 2024) The optimal hyperparameter parameter values based different ML algorithms for hd model. Table 4The modelling results of different ML algorithms.AlgorithmTrainValidationR2RMSEMSEMAER2RMSEMSEMAECatBoost0.86072.78697.76692.12400.76803.609513.02842.6681LGBM0.97051.28251.64470.96630.79013.433511.78862.4794XGBoost0.89022.47426.12141.91830.77593.547212.58242.6148Note: MAE, mean absolute error (m); MSE, mean squared error (m²); RMSE, root mean squared error (m); R², coefficient of determination. The modelling results of different ML algorithms. Note: MAE, mean absolute error (m); MSE, mean squared error (m²); RMSE, root mean squared error (m); R², coefficient of determination. To account for ecological heterogeneity across vegetation divisions, the MLME model was constructed by integrating LGBM predictions (fixed effects) with vegetation division (random effects). Two random-effect structures, (1 | zone) and (fixed_hd | zone), were evaluated using AIC and BIC (Table 5). The inclusion of random slopes significantly improved model performance (p < 0.001). Vegetation zone–specific parameter estimates (Table 6) revealed spatially varying model sensitivities, with higher slope adjustments in temperate desert (TD) warm temperate deciduous-broadleaf forest (WT) and cold temperate (CT) zones, indicating distinct forest structure responses. Table 5The fitting results of ML-based mixed-effects models.EquationNparAICBIClogLikChisqPr( > Chisq)hd~fixed_hd+(1|zone)427085692708614-13542812708561hd~fixed_hd+(fixed_hd|zone)627080962708164-13540422708084 < 2.2e-16***Note: (1 | zone) specifies random intercepts for the grouping factor zone;while (x | zone) specifies zone-specific random slopes for the covariate; Npar is the number of estimated parameters; The symbol “***” denotes a p-value less than 0.001. The fitting results of ML-based mixed-effects models. Note: (1 | zone) specifies random intercepts for the grouping factor zone;while (x | zone) specifies zone-specific random slopes for the covariate; Npar is the number of estimated parameters; The symbol “***” denotes a p-value less than 0.001. Table 6Parameter estimates of MLME models for vegetation zones.Vegetation zonebβCold temperate needleleaf forest−0.9921.055Temperate needleleaf-broadleaf mixed forest−0.8851.059Tropical monsoon forest-rainforest−0.6051.023Temperate steppe−0.8601.045Temperate desert−8.8211.201Warm temperate deciduous-broadleaf forest−1.0641.053Subtropical evergreen broadleaf forest−0.7051.049 Parameter estimates of MLME models for vegetation zones. The MLME framework achieved robust accuracy across China’s forests, with R² = 0.7907, RMSE = 3.4249 m, and MAE = 2.4733 m (Table 7). Compared with standalone ML models, MLME showed slightly improved accuracy, confirming that incorporating vegetation divisions enhances model generalizability and ecological interpretability. The agreement between the MLME estimates and the LiDAR validation observations is displayed in Fig. 2. A strong agreement is observed between the estimated and validation results, with a high correlation in the MLME-based hd model (r = 0.890). Overall, these results demonstrate that the MLME-based hd estimates possess high technical quality, spatial consistency, and strong predictive reliability for large-scale forest monitoring. Table 7Validation metrics for different hd modeling approaches (MLs and MLME).AlgorithmR2RMSEMSEMAECatBoost0.76803.609513.02842.6681LGBM0.79013.433511.78862.4794XGBoost0.77593.547212.58242.6148MLME_LGBM0.79073.424911.72992.4733Note: MAE, mean absolute error (m); MSE, mean squared error (m²); RMSE, root mean squared error (m); R², coefficient of determination. Validation metrics for different hd modeling approaches (MLs and MLME). Note: MAE, mean absolute error (m); MSE, mean squared error (m²); RMSE, root mean squared error (m); R², coefficient of determination. Scatter plot comparing the MLME-derived hd with LiDAR validation data. Forest dominant height dataset for China The nationwide hd dataset revealed substantial spatial heterogeneity across China’s vegetation divisions (Fig. 3). The maximum hd reached 101 m in the subtropical evergreen broadleaf forest, followed by the tropical monsoon forest–rainforest (77 m) and temperate desert (52 m). Moderate values were found in the Qinghai–Tibet Plateau alpine vegetation (44 m), temperate steppe (39 m), temperate needleleaf–broadleaf mixed forest (36 m), cold temperate needleleaf forest (34 m), and warm temperate deciduous broadleaf forest (32 m) (Fig. 3, Table 8). In contrast, mean hd exhibited distinct patterns (Table 8). The tropical monsoon forest–rainforest had the highest mean hd (18.57 ± 5.62 m), followed by the subtropical evergreen broadleaf forest (15.71 ± 3.75 m), despite the latter hosting the tallest trees. The Qinghai–Tibet Plateau alpine vegetation recorded the lowest mean hd (12.34 ± 4.11 m). Forest dominant height across China at 30 m spatial resolution in 2020. Table 8Forest dominant height analysis across different vegetation divisions.Vegetation divisionMean(m)SD(m)Median(m)Tropical monsoon forest-rainforest18.575.6218.00Temperate needleleaf-broadleaf mixed forest17.413.3718.00Temperate steppe16.573.8417.00Cold temperate needleleaf forest16.512.6417.00Subtropical evergreen broadleaf forest15.713.7515.00Temperate desert14.277.4613.00Warm temperate deciduous broadleaf forest14.022.6614.00Qinghai-Tibet Plateau alpine vegetation12.354.1113.00 Forest dominant height analysis across different vegetation divisions. Uncertainty analysis As shown in Fig. 4, the uncertainty associated with hd at the 30 m pixel scale was assessed by aggregating three error components: measurement, product, and prediction errors. The overall uncertainty for mapping hd at the pixel scale stood at 22.73%. Nevertheless, the spatial distribution of uncertainty varied significantly across China, driven by regional forest heterogeneity, the quality of remote sensing inputs, and the spatial coverage and sample size of the reference data (UAV LiDAR data). Estimations across the country showed a variation of no more than 1%, indicating strong spatial consistency and robust model performance at the national scale. Spatial distribution of pixel-level uncertainty in hd. Vegetation divisions with sparse LiDAR coverage, such as Temperate desert and Cold temperate needleleaf forest, tended to show higher uncertainty (Table 1 and Supplementary Table S4). By contrast, Temperate steppe, which had the highest sampling proportion, showed comparatively lower uncertainty. In addition, vegetation divisions with more complex forest structure, such as Subtropical evergreen broadleaf forest and Tropical monsoon forest–rainforest, still showed moderate uncertainty despite relatively better sampling coverage, indicating that stand heterogeneity and vertical structural complexity may also contribute to uncertainty. Qinghai–Tibet Plateau alpine vegetation also showed elevated uncertainty (Supplementary Table S4). Because this division was not represented in the UAV LiDAR coverage statistics in Table 1, its elevated uncertainty is more likely associated with regional environmental complexity, such as terrain heterogeneity. Usage Notes The national wall-to-wall forest dominant height dataset provides substantial value for large-scale forest resource assessment and ecological research. hd is a key stand-level parameter that reflects site quality and long-term productivity potential. Because it is derived from the tallest trees within a stand, hd is largely independent of stand density, thinning intensity, or understory competition, and therefore exhibits stable age–height trajectories across diverse management regimes9,13,39,40. Owing to these properties, hd constitutes a fundamental variable for site index determination and is widely used in growth and yield modelling, biomass estimation, and a range of operational forest management applications. At the national scale, continuous, wall-to-wall hd products derived from individual-tree detection remain extremely rare. To date, such datasets exist (Table S3) only in a handful of countries with near-complete airborne LiDAR coverage (e.g., Finland41, parts of Canada42, and selected U.S. regions43, and even there they are typically restricted to managed forests, specific species, or inconsistent methodologies that deviate from national forest inventory definitions. The present dataset advances beyond existing products by delivering the first nationwide 30 m hd map in China derived directly from individual-tree segmentation and detection using close-range UAV LiDAR data. Its key advantages include: Data acquisition: Levering an unprecedented collection of > 1,117 km² of high-density UAV LiDAR data systematically acquired across all eight major vegetation divisions of the country, providing a nationally representative training and validation basis under diverse stand structures and species assemblages. Data acquisition: Levering an unprecedented collection of > 1,117 km² of high-density UAV LiDAR data systematically acquired across all eight major vegetation divisions of the country, providing a nationally representative training and validation basis under diverse stand structures and species assemblages. Processing sophistication: Application of both ecozone-specific models and a unified national mixed-effects modeling framework, which balances local model adaptability with cross-ecozone consistency and minimizes artificial discontinuities at ecozone boundaries. Processing sophistication: Application of both ecozone-specific models and a unified national mixed-effects modeling framework, which balances local model adaptability with cross-ecozone consistency and minimizes artificial discontinuities at ecozone boundaries. Despite these strengths, several limitations should be acknowledged to guide appropriate data use and future improvements. Its limitations include: The spatial distribution of UAV LiDAR data is uneven at the national scale, with relatively sparse coverage in regions such as the Qinghai–Tibet Plateau and parts of northwestern China. Coverage is also limited in some ecosystem types, particularly Temperate desert and Cold temperate needleleaf forest (Table 1). Although the mixed-effects modeling framework partially mitigates this imbalance by borrowing strength across ecozones, localized uncertainties may persist in under-sampled regions. Accordingly, dominant height estimates in these under-sampled regions should be interpreted with caution, and expanded UAV LiDAR sampling in these areas will be important for improving future versions of the dataset. The spatial distribution of UAV LiDAR data is uneven at the national scale, with relatively sparse coverage in regions such as the Qinghai–Tibet Plateau and parts of northwestern China. Coverage is also limited in some ecosystem types, particularly Temperate desert and Cold temperate needleleaf forest (Table 1). Although the mixed-effects modeling framework partially mitigates this imbalance by borrowing strength across ecozones, localized uncertainties may persist in under-sampled regions. Accordingly, dominant height estimates in these under-sampled regions should be interpreted with caution, and expanded UAV LiDAR sampling in these areas will be important for improving future versions of the dataset. The LiDAR datasets were acquired over a multi-year period using different sensors, which may introduce residual inconsistencies despite careful preprocessing and calibration. Temporal mismatches related to tree growth cannot be fully eliminated under current data availability but represent a common constraint in large-scale tree-based height mapping efforts. The LiDAR datasets were acquired over a multi-year period using different sensors, which may introduce residual inconsistencies despite careful preprocessing and calibration. Temporal mismatches related to tree growth cannot be fully eliminated under current data availability but represent a common constraint in large-scale tree-based height mapping efforts. Uncertainties associated with individual tree segmentation, including omission and commission errors in structurally complex stands, may propagate into hd estimates, although these effects are explicitly accounted for in the uncertainty analysis presented in this study. Uncertainties associated with individual tree segmentation, including omission and commission errors in structurally complex stands, may propagate into hd estimates, although these effects are explicitly accounted for in the uncertainty analysis presented in this study. An additional limitation is the lack of explicit tree species information for the tallest trees used to derive hd. As a result, species-specific dominance relationships cannot be distinguished in mixed-species stands, which may influence interpretations where hd is linked to species-specific growth potential. Nevertheless, the dataset remains fully consistent with widely adopted inventory-based definitions of hd that rely on height ranking rather than species identity. An additional limitation is the lack of explicit tree species information for the tallest trees used to derive hd. As a result, species-specific dominance relationships cannot be distinguished in mixed-species stands, which may influence interpretations where hd is linked to species-specific growth potential. Nevertheless, the dataset remains fully consistent with widely adopted inventory-based definitions of hd that rely on height ranking rather than species identity. Overall, this dataset unlocks a suite of previously unattainable national-scale applications, including site productivity and site index mapping, parameterization of regional and national growth and yield systems, improved forest biomass and carbon accounting, nationwide assessment of vertical forest structure, and robust benchmarking of current and future spaceborne LiDAR missions (e.g., GEDI, MOLI). By providing a standardized, reproducible hd product, the dataset offers a foundational data resource for forest structure analysis and model development at large spatial scales. Supplementary information Supplementary information Acknowledgements The authors acknowledge the collective efforts of all contributors involved in the data collection and preparation processes. Author contributions Y.L.C., H.T.Y., G.C.X., and Q.H.G. conceived and designed the project. Y.L.C. and H.T.Y., and G.C.X. performed the dominant tree height mapping, validation, and analysis. H.T.Y. and Y.L.C. wrote the first draft together. G.C.X. conducted field sampling. All authors contributed to the results and revision of the manuscript. Funding This research was financially supported by the International Research Centre of Big Data for Sustainable Development Goals [CBAS2022GSP06], the National Key Research and Development Program [2022YFF13002003 and 2022YFF1300202], and the National Natural Science Foundation of China [42371329 and 32301285]. Data availability The forest dominant height dataset generated in this study is publicly available at Zenodo: https://doi.org/10.5281/zenodo.1828606938. All input datasets used in this study are publicly available from the original data providers and are described in the Methods. Sentinel-1 SAR GRD data were accessed from the Google Earth Engine Data Catalog as collection COPERNICUS/S1_GRD (https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD). Landsat 8 Collection 2 Level-2 data were accessed from the Google Earth Engine Data Catalog (https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LC08_C02_T1_L2). Climate data were obtained from WorldClim (https://worldclim.org/data/index.html). Soil data were obtained from the Harmonized World Soil Database version 1.2 (HWSD v1.2; https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/). SRTM elevation data were obtained from the USGS EarthExplorer portal (https://earthexplorer.usgs.gov). The China forest age product used in this study is publicly available at Zenodo: https://doi.org/10.5281/zenodo.8354262. Code availability The code is publicly accessible on Github: ChenYL2021/ForestDominantHeightDataset: Mapping forest dominant height of China's forests by UAV-LiDAR data. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Yuling Chen, Guangcai Xu. Contributor Information Haitao Yang, Email: yht90h@163.com. Qinghua Guo, Email: guo.qinghua@pku.edu.cn. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41597-026-07546-z. References Associated Data Data Citations Supplementary Materials Supplementary information Data Availability Statement The forest dominant height dataset generated in this study is publicly available at Zenodo: https://doi.org/10.5281/zenodo.1828606938. All input datasets used in this study are publicly available from the original data providers and are described in the Methods. Sentinel-1 SAR GRD data were accessed from the Google Earth Engine Data Catalog as collection COPERNICUS/S1_GRD (https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD). Landsat 8 Collection 2 Level-2 data were accessed from the Google Earth Engine Data Catalog (https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LC08_C02_T1_L2). Climate data were obtained from WorldClim (https://worldclim.org/data/index.html). Soil data were obtained from the Harmonized World Soil Database version 1.2 (HWSD v1.2; https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/). SRTM elevation data were obtained from the USGS EarthExplorer portal (https://earthexplorer.usgs.gov). The China forest age product used in this study is publicly available at Zenodo: https://doi.org/10.5281/zenodo.8354262. The code is publicly accessible on Github: ChenYL2021/ForestDominantHeightDataset: Mapping forest dominant height of China's forests by UAV-LiDAR data.