To assess the quantity, structure, function, and productivity of national forest resources, the National Forestry and Grassland Administration of China conducted ten national forest resource inventories between 1973 and 2023. The inventories took place during the periods 1973–1976, 1977–1981, 1984–1988, 1989–1993, 1994–1998, 1999–2003, 2004–2008, 2009–2013, 2014–2018, and 2019–2023. The data are available from the National Forestry and Grassland Science Data Center (NFGSDC) at https://www.forestdata.cn/ (last access: 20 June 2025). The surveys were performed at the provincial level, and a systematic sampling design with fixed plots located at the intersections of the national 1:50000 or 1:100000 topographic map grids were employed. For each plot, the recorded variables included forest cover, forest type area, and standing volume. Provincial-level area statistics from NFI report specifically, area data for wooded land, needleleaf and broadleaf forests from the second through eighth inventories; arbor forestland area from the ninth inventory; and total forestland area from the sixth through tenth inventories, were used in the analyses.

In this study, twenty-two datasets covering the period 1981–2023 were utilised as the foundational inputs for forest cover reconstruction (Table S1). Forest cover information was extracted from these LULC products. Preprocessing of the data involved several steps: (i) reprojecting all the datasets to the WGS_1984_Albers spatial reference system; (ii) resampling to a 1 km resolution using the majority method; and (iii) aligning all the data to a common grid framework to ensure a consistent cell size and spatial extent for China.

In this study, China's first seamless annual leaf-on (growing season) Landsat composite dataset (1985–2023) from Cai et al. (2025) was utilised. This dataset harmonises multisensor Landsat imagery through a comprehensive compositing method, addressing critical issues such as cloud/shadow contamination, reflectance consistency, and data gaps, thereby providing a single annual growing-season composite image. The dataset is available on the Google Earth Engine (GEE) platform (https://ee-caiyt33-catcd.projects.earthengine.app/view/landsat-yearly-composite-viewer, last access: 20 October 2025).

From this dataset, we derived annual growing-season NDVI data using the near-infrared (NIR) and red bands, calculated as follows: (NIR-Red)/(NIR+Red). These data were then resampled to a 1 km resolution using mean aggregation. In this study, the growing-season NDVI served as the primary indicator for classifying “potential forestland pixels” as forestland (see Sect. 3.3). Owing to the scarcity of Landsat imagery in the early 1980s (Yang and Huang, 2021a) and to ensure temporal consistency across the extended 1981–2023 analysis period, records for 1981–1984 were gap-filled using data from 1985. To validate this substitution, we performed a pixel wise correlation analysis between the 1985 growing season NDVI and three independent, coarser resolution products covering 1981–1985 (Jeong et al., 2024; Li et al., 2023; Li et al., 2024), with all datasets processed as growing season (June–October) median composites and the 1985 baseline resampled to match the respective coarse resolutions. The results indicate that the 1985 dataset maintained high spatial consistency with the validation products (R2 > 0.8). Furthermore, the correlation metrics exhibited exceptional temporal stability (ΔR2 < 0.04; Fig. S1 in the Supplement).

To assign phenological types to the small number of remaining unclassified forestland pixels (see Sect. 3.4.2), two supplementary regional partitioning products were utilised. The first was the Köppen–Geiger climate classification from Beck et al. (2018), which classifies Earth's land surface into 30 distinct climate zones at a 0.0083° resolution (approx. 1 km) on the basis of temperature and precipitation records from 1980–2016. The data are publicly available from Figshare at https://figshare.com/articles/dataset/Present_and_future_K_ppen-Geiger_climate_classification_maps_at_1-km_resolution/6396959/2 (last access: 26 May 2025). The second product was a global topographic dataset from Amatulli et al. (2018), derived from the 250 m GMTED2010 and 90 m SRTM4.1dev digital elevation models. In this dataset, the global land surface is classified into ten topographic categories: flat, peak, ridge, shoulder, spur, slope, hollow, footslope, valley, and pit. The data are publicly available from Earthenv at https://www.earthenv.org/topography (last access: 25 May 2025). In general, needleleaf forests are predominantly evergreen, with the notable exception of larch forests, which are deciduous and found mainly in boreal regions. In contrast, broadleaf forests are typically deciduous, although those in tropical regions are predominantly evergreen. To further classify evergreen and deciduous forest types as either boreal/temperate or tropical (see Sect. 3.4.3), the 1 km resolution monthly mean temperature dataset for China (1981–2023) from Peng et al. (2019a) was utilised. This dataset was generated by spatially downscaling the global 0.5° CRU climate dataset and the high-resolution WorldClim dataset using the delta spatial downscaling method. The data are publicly available from the National Tibetan Plateau Data Center (10.11888/Meteoro.tpdc.270961, Peng, 2019).

A clear definition of forest composition is needed prior to reconstructing the PFT distribution. In accordance with China's “Technical Regulations for Continuous Forest Inventory” (GB/T 38590–2020), China's forests are divided into categories such as wooded land, arbor forestland, shrubland, bamboo forestland, and economic forestland, following a specific hierarchical classification system (Table 2). The area relationships among these components are defined as follows in Eqs. (1)–(3): 1AForest=AWoodedland+AShrubland2AWoodedland=AArbor+ABamboo+AEcon3AArbor=ANeedleleaf+ABroadleaf.

Our reconstruction method adheres to two primary constraints: (1) Forest area constraint: To ensure consistency between statistics and remote sensing data, our “forest” class is constrained using the provincial-level area of wooded land. We explicitly exclude the shrubland category. This approach aligns our statistical constraint with the “forest” class (i.e., tree-covered areas) typically identified in LULC products, excluding the “shrub” category present in some LULC products. (2) PFT area constraint: The subclassification of PFTs is constrained using the provincial area of arbor forest, which comprises the needleleaf and broadleaf categories (Table 2). Notably, the NFI technical regulations specify that “mixed leaf forest” is included within the needleleaf forest area. This two-level constraint is used to determine the mapping approach used to identify potential forestland pixels by harmonising corresponding forestland classes from diverse LULC products, while ensuring that the reconstructed forestland area and the areas of our derived PFTs are strictly consistent with their corresponding NFI statistical areas.

To construct the continuous forest cover time series for 1981–2023, provincial forest areas from each NFI period were assigned to the final year of that period (e.g., data from the 1976–1981 survey were assigned to 1981) (Yue et al., 2024). However, owing to the latency in NFI data publication, the temporal coverage of statistical data varies by category. Notably, the statistics for wooded land, needleleaf forests, and broadleaf forests are available only up to the eighth inventory (2013); arbor forest statistics are available up to the ninth inventory (2018); and the total forest area is available up to the tenth inventory (2023). To address these data gaps, we extrapolated the missing values for the post-2013 period using linear regression models based on published data. This approach is supported by prior studies, which indicated that national afforestation targets exhibit a consistent linear trend over time, increasing by approximately 1.8 million ha annually (He et al., 2024a; Xu et al., 2023).

Specifically, we first modelled the relationship between provincial total forest area and arbor forest area using available data from 2003–2018. We applied these models to extrapolate the arbor forest area for 2023, using the known 2023 total forest area as the predictor (Fig. S2; China R2 = 0.97). Concurrently, using data from 2003–2013, we established linear regression models between total forestland area and wooded land area to preliminarily predict wooded land areas for 2018 and 2023. According to the definition in Sect. 3.1, the area of wooded land must be strictly equal to or exceed that of arbor forest. However, a comparison of our preliminary wooded land predictions from 2018 with the recorded arbor forest data from 2018 revealed logical inconsistencies in Tibet and Gansu Provinces, where the predicted wooded land area was smaller than the arbor forestland area. To rectify this difference, we adjusted the 2018 wooded land values for these two provinces to match their actual arbor forestland areas and reincorporated these corrected data into the model to predict the 2023 wooded land area. A final consistency check of the 2023 predictions revealed a similar discrepancy in Qinghai Province, which was adjusted accordingly to match the predicted arbor forest area (Fig. S3; China R2 = 0.98).

Next, to estimate the needleleaf and broadleaf forestland areas for 2018 and 2023, we assumed that the growth trends observed from 2003–2013 persisted into the 2013–2023 period (Chini et al., 2021). Given that the arbor forestland area is the sum of the needleleaf and broadleaf forestland areas (see Sect. 3.1), we calculated the ratio of each forest type to the total arbor forestland area for 2003, 2008, and 2013. We then constructed linear regression models based on these historical ratios to project the corresponding proportions for 2018 and 2023. The final provincial areas for needleleaf and broadleaf forests were derived by multiplying these projected ratios by the (known or extrapolated) arbor forest areas for 2018 and 2023. Finally, once the areas for all benchmark years were established, we applied linear interpolation to fill the gaps in annual data between these benchmarks (Yu et al., 2022; Yue et al., 2024). All extrapolation and interpolation procedures were implemented independently for each province to account for regional heterogeneity in policies and environmental factors. Notably, arbor forest statistics are missing for Hong Kong, Macau, and post-2013 Taiwan. Given that China's major afforestation programs do not cover these regions, we did not reconstruct the arbor forest areas for Hong Kong, Macau, or post-2013 Taiwan (Zhang et al., 2025).

In this study, a data-driven “forest consistency” method was developed to reconstruct historical forest cover (Fig. 1). The method involved overlaying all available LULC datasets for each year (Fig. S4). For any given pixel, “consistency” (CON) was defined as the number of datasets in which the pixel was classified as forestland (Fig. S5a). A pixel was subsequently identified as a “potential forest pixel” if it was classified as forest in at least one dataset (i.e., CON > 0). The consistency value was then used to establish priority, whereby a high CON value indicated a high likelihood of the pixel representing true forest cover (Xia et al., 2023; Fang et al., 2020).

To determine the consistency threshold for the final forest classification, all potential forestland pixels were ranked in descending order of their CON values. The NFI-derived area for a given province was used as the target value to establish this threshold. Specifically, two scenarios were considered. First, if the total area of all potential forestland pixels was less than the NFI-reported area, all potential pixels were classified as forestland, a scenario that could result in an underestimation for that province. Second, if the total area of potential pixels exceeded the NFI-reported area, a cumulative summation was performed. Pixels were incrementally summed, starting from the highest consistency value downwards, until the cumulative area bracketed the NFI target area (ANFI). If the cumulative area of pixels with CON≥m was less than ANFI but the cumulative area of pixels with CON≥m-1 exceeded ANFI, then m was defined as the consistency threshold. All pixels with a consistency value ≥m were subsequently classified as forestland. To precisely match the NFI target area, however, a portion of the remaining required area was filled by selecting pixels from the CON=m-1 level. On the basis of the assumption that, within a given consistency level, a high NDVI value indicates a high likelihood of forest cover existing, the growing-season NDVI was calculated for all pixels at the CON=m-1 level. These pixels were then ranked in descending order of their NDVI values. The top n pixels were subsequently selected as “residual forest pixels”, where n was determined on the basis of the remaining area required to precisely match the NFI target.

In summary, the final forest classification was performed to identify pixels that exhibited both high cross-dataset consistency and high growing-season NDVI values. The total area of this final classification was strictly constrained by provincial NFI statistics, thereby ensuring that the reconstructed maps aligned with the authoritative inventory data. While this method generally ensures a close correspondence to the NFI-reported area, minor systematic underestimation can occur. This is a consequence of pixel resolution limitations, particularly when the final area required to meet the NFI target is smaller than that of a single pixel.

To validate the accuracy of our reconstructed annual PFT maps, we collected an extensive validation dataset covering the entire study area. Our validation samples were labelled on the basis of three data sources: (1) ground plots from the NFI conducted between 2009 and 2013. For each plot, the dataset provides the plot ID, geographic coordinates, and a classification into one of five forest types: ENF, EBF, DNF, DBF, and MF. (2) A collection of Landsat images spanning 1985–2023 (Cai et al., 2025) and (3) high-resolution imagery from Google Earth. First, to assess the temporal consistency of the NFI ground plots, we applied the normalised burn ratio (NBR) (García and Caselles, 1991) to the dense Landsat time series at these plot locations. The NBR is calculated as follows: (NIR-SWIR)/(NIR+SWIR). This index is recognised for its sensitivity in detecting forest disturbances within Landsat time series (Perbet et al., 2024; White et al., 2022). We then employed the Landsat-based detection of trends in disturbance and recovery (LandTrendr) algorithm (Kennedy et al., 2010) to assess the stability of the 1 km grid cells surrounding these nationwide plots from 1985 to 2023. LandTrendr is a widely used disturbance detection method designed to detect anomalies and trends in time series, distinguishing between abrupt and gradual changes (Kennedy et al., 2018; Cheng et al., 2024). We implemented this algorithm in the GEE to distinguish between stable (remained unchanged between 1985 and 2023) and unstable (underwent a change in at least one year from 1985–2023) samples among the NFI ground plots and to identify the timing of any detected changes. This process yielded 5481 stable and 1229 unstable plots.

For unstable plots, we employed a cross-validation strategy to pinpoint the precise timing of land cover transitions. Specifically, we integrated raw NBR values derived from Landsat time series, LandTrendr-fitted temporal trajectories, and historical high-resolution imagery from Google Earth. We paid particular attention to validating the image features corresponding to the segments with the greatest magnitude of change between fitted vertices (Figs. S8 and S9). Through visual interpretation of spectral variations and textural features in the Google Earth imagery, we identified the exact year of land cover conversion (e.g., forestland converted to nonforest in 2013; Fig. S8). Consequently, the specific PFT label for each plot was consistently applied to all years within the 1985–2023 period where the plot was identified as forestland.

For the 1981–1984 period, we extrapolated the classification status from 1985 to the preceding years (i.e., plots identified as a specific forest PFT or nonforest in 1985 were assigned the same attribute for 1981–1984). This approach assumes that reversals between forest and nonforest states (e.g., forest–nonforest–forest transitions) are unlikely to occur within a 5 year window. The labelling was conducted by ten experts experienced in remote sensing data analysis. To ensure rigorous quality control, the samples with ambiguous features were subjected to team discussion, and the final reference labels were assigned only after a majority consensus. Finally, by integrating the annual records from both stable and unstable plots, we constructed a validation database spanning the entire 1981–2023 period.

Following the good practice guidelines for sample size decisions proposed by Olofsson et al. (2014), we used a proportional stratified sampling design to randomly draw validation samples from the validation database for each year, stratified on the basis of the five PFT classes: ENF, EBF, DNF, DBF, and MF. Each annual validation set consisted of approximately 1280 samples. For the proportionally rare classes (DNF and MF), the sample size was increased to 100 samples (Fig. S10) to reduce the standard error of the accuracy estimates for these less common categories. The PFT class of each sample was further confirmed following the same visual interpretation protocol described above. On the basis of these validation samples, we generated a confusion matrix for each year and calculated the producer's accuracy (PA), user's accuracy (UA), overall accuracy (OA), and F1-score: 5PAk=pkk∑j=1mpkj6UAk=pkk∑i=1mpik7OA=∑k=1mpkk 8F1=2×PA×UAPA+UA×100% where pkk is the area proportion of class k that is correctly classified (i.e., class k in both the reference data and the map); ∑j=1mpkj represents the total area proportion of class k in the reference data (the row sum), ∑i=1mpik represents the total area proportion of class k in the classified map (the column sum), and m is the number of land cover types. We then corrected the accuracy estimates on the basis of map uncertainties and calculated the corresponding standard errors and 95 % confidence intervals (Olofsson et al., 2014). We also conducted cross-validation with existing annual land cover products (ESA CCI, MCD12Q1, and GLC_FCS30D), which included five PFT classes (ENF, EBF, DNF, DBF, and MF) to comprehensively evaluate the quality of the dataset reconstructed in this study.

In addition to this sample-based validation, an indirect accuracy assessment was performed by analysing the consistency among the input LULC datasets. This approach is pertinent as the final product is an integration of information from these sources. Here, consistency is defined at the pixel level as the number of LULC datasets with the same classification of a specific forest type. The underlying assumption of this analysis is that a high consistency value for a given pixel indicates high confidence in its classification and a high likelihood of it being correct (Xia et al., 2023).

Forest change is defined as the transition in land cover between forestland and nonforest states over a given period. This type of change is typically classified into change events (i.e., forest gain or loss) and stable states (i.e., persistent forest or persistent nonforest) (Winkler et al., 2021). Forest gain represents a transition from a nonforest state to a forest state, whereas forest loss is the reverse process (Hansen et al., 2013). To identify the onset year and duration of forest change events across China for the period 1981–2023, a pixel-level time series analysis was developed on the basis of the annual forest mask sequence.

This methodology is illustrated here using the detection of forest gain. First, the annual forest masks were standardised to binary values (0 = nonforest; 1 = forest) to create a spatiotemporal data cube. For this analysis, a “stable” forest state was defined as a pixel remaining as forestland for at least five consecutive years. The onset year of a forest gain event was then identified for each pixel as the first year it transitioned from a nonforest state to a stable forest state.

Following the identification of a gain event, the duration of forest persistence was calculated. This duration is the number of years from the onset of the gain until either (a) the pixel underwent a stable loss event, defined as transitioning to nonforest and remaining so for at least five consecutive years, or (b) the end of the study period (2023) if no such loss event occurred. The detection of forest loss events and their duration was based on the inverse logic.

This analysis produced four maps: two indicating the onset year for forest gain and loss events and two representing the durations of these respective periods.

The impact of the new PFT distribution on surface fluxes was assessed using LPJ–GUESS (Lindeskog et al., 2021), a process-based DGVM. The model was driven with Chinese meteorological forcing data (CMFD 2.0) (He et al., 2020) at a 0.1° resolution for the period of 1951–2023, this dataset included daily averages of temperature, incoming shortwave radiation, and precipitation. The data are publicly available from the National Tibetan Plateau Data Center at 10.11888/Atmos.tpdc.302088 (He et al., 2024b). Atmospheric CO₂ concentrations followed historical trajectories (Friedlingstein et al., 2025). Gridded monthly nitrogen deposition data (0.5° resolution) were also used as inputs (Lamarque et al., 2013), using the nearest grid cell value for each simulation grid. Soil properties – namely, the fractions of sand, silt, and clay; organic carbon content; C:N ratio; pH; bulk density; and organic carbon density – were derived from the China dataset of soil properties for land surface modelling version 2 (CSDL v2) dataset (Shi et al., 2025). This database, generated using advanced integrated machine learning algorithms from multiple representative historical soil profiles and high-resolution environmental covariates, provides 0.1° maps of six soil layers across China. The data are publicly available at https://www.scidb.cn/s/ZZJzAz (last access: 23 September 2025). We averaged the soil texture data from 0–200 cm and resampled all the soil inputs to align them with the meteorological forcing grid.

The LPJ-GUESS model was run at a 0.1° × 0.1° spatial resolution. Vegetation structure and its associated carbon, water, and nitrogen pools were initialised using a 500 year spin-up phase (starting from bare ground) by cycling 1951–1980 detrended meteorological data. This was followed by a historical simulation from 1981–2023. To account for natural disturbances, a disturbance interval of 100 years was established, and the GlobFirm wildfire submodel (Thonicke et al., 2001) was used. To ensure the ability of the model to simulate Chinese forest ecosystems, key parameters related to photosynthesis, autotrophic respiration, and plant water use efficiency were manually adjusted after benchmarking (all the updated PFT parameters are summarized in Table S2).

To isolate and quantify the independent impact of the PFT distribution, we designed three distinct model experiments. During the 1981–2023 period, all three experiments were driven by the same climate forcing data and consistent model parameters. The only difference among the experiments was the PFT input map: EXP1 utilised the new PFT map developed in this study, EXP2 utilised the global PFT dataset from ESA CCI (Harper et al., 2023), and EXP3 utilised the global PFT dataset from MCD12Q1 (Sulla-Menashe et al., 2019). In all three experiments, we cycled the 2010 PFT distribution to maintain a static land cover driver. We simulated the monthly total GPP, NEE, LAI, and ET for 2010, focusing our comparison on the mean differences between the PFT datasets during the summer (June–August) to highlight primary changes. As the PFT distribution map was the sole variable altered between experiments, any observed differences in the output fluxes could be directly attributed to its influence. To assess the plausibility of these changes, the simulation results were also compared to four remote sensing and observation-based products: FLUXCOM GPP, FLUXCOM NEE, GIMMS LAI4g, and GLEAM ET. All the observational products were resampled to a 0.1° resolution using bilinear interpolation to match the LPJ–GUESS simulation grid.

Independent validation based on five forest types for the period 1981–2023 demonstrates that the accuracy of the PFT dataset reconstructed in this study is both stable and satisfactory. The OA ranged from 79.74 % to 88.5 %, with a mean OA of 84.86 % ± 1.18 % (Fig. 2f). Among the specific forest classes, the highest mean F1 score was achieved for EBF (93.03 %), followed by the ENF (92.49 %), DBF (84.89 %), DNF (62.29 %), and MF (54.03 %) classes (Fig. 2a–e). Furthermore, in most years, the overall accuracy of our dataset (mean 84.86 % ± 1.18 %) was greater than that of ESA CCI (mean 83.47 % ± 1.15 %), MCD12Q1 (mean 61.17 % ± 1.36 %), and the global 30 m land-cover dynamics monitoring dataset (GLC_FCS30D) (mean 78.92 % ± 1.24 %) (Fig. 2f). For the classes with large areal proportions, such as DBF, EBF, and ENF, the F1 scores for our dataset also demonstrated superior and more stable performance than ESA CCI, MCD12Q1, and GLC_FCS30D did (Fig. 2a–e). This independent validation confirms the high spatial accuracy of the reconstructed PFT data in comparison to prominent existing datasets. Moreover, the temporal coverage of the dataset reconstructed in this study spans 43 years (1981–2023), exceeding that of ESA CCI (1992–2020), MCD12Q1 (2001–2023), and GLC_FCS30D (1985–2022). Analysis of the consistency between our reconstructed PFT dataset and the input LULC datasets revealed that the reconstructed broadleaf forests were more consistent than the needleleaf forests were (Fig. S5b–f), indicating greater confidence in the reconstructed BF results. Among the five PFTs, the confidence ranking from highest to lowest was DBF, EBF, ENF, DNF, and MF (Fig. S5b–f). Furthermore, according to our methodology, a pixel-weighted average over the 1981–2023 period indicates that fewer than 1 % of the pixels for all five reconstructed PFTs did not fall within their corresponding consistency type classes (the specific proportions were 0.50 % for DBF, 0.10 % for EBF, 0.03 % for DNF, 0.01 % for ENF, and 0.05 % for MF, Fig. S11).

The internal consistency of the reconstructed total forest cover was assessed at ten specific time points: 1981, 1985, 1990, 1995, 2000, 2005, 2010, 2015, 2020, and 2023 (Fig. 3). The number of inputs LULC datasets available for reconstruction varied at each time point, with 5, 7, 6, 7, 11, 11, 11, 10, 12, and 6 products used for each respective year. The analysis revealed that for each of the ten time points, the maximum possible consistency score was achieved for 95.7 %, 93.5 %, 91.8 %, 85.3 %, 66.7 %, 54.9 %, 55.7 %, 57.9 %, 55.4 %, and 85.4 % of the reconstructed forestland pixels, respectively. Conversely, pixels with the lowest possible consistency (CON = 1) consistently accounted for a small fraction of the total reconstructed forest area, ranging from 0 % to 0.26 % across different years (Fig. 3). Spatially, areas with lower forest consistency were predominantly located in the arid and semiarid regions of northwestern China (e.g., Ningxia) and the highly fragmented landscapes of the eastern coastal plains (e.g., Tianjin, Shandong, Jiangsu, and Shanghai). In contrast, high-consistency forestland areas were concentrated mainly in regions with extensive and stable forest cover, primarily in southern and central China, including provinces such as Hubei, Zhejiang, Guangxi, Guizhou, Yunnan, and Jiangxi (Fig. S12). A comparison and an analysis of the NFI data with the forest area estimates reconstructed in this study at the provincial scale (Fig. S13) reveal a good match for multiple years (1981–2023), with an R2 close to 1 and a p-value < 0.001. These findings indicate that the reconstructed data from this study are highly consistent with the NFI statistics in terms of overall trends.

This dataset provides the annual forest cover distribution from 1981 to 2023 (Fig. 3) and the distribution of nine PFTs at a 1 km spatial resolution (Fig. 4). The data are supplied in the WGS 1984 Albers equal-area conic projection. The nine PFTs are (1) boreal evergreen needleleaf forest, (2) temperate evergreen needleleaf forest, (3) temperate evergreen broadleaf forest, (4) boreal deciduous broadleaf forest, (5) temperate deciduous broadleaf forest, (6) tropical deciduous broadleaf forest, (7) tropical evergreen broadleaf forest, (8) deciduous needleleaf forest, and (9) mixed leaf forest.

For the reference year 2023, the dataset indicates that China's forests are composed of temperate evergreen broadleaf forest (27.5 %), temperate evergreen needleleaf forest (23.4 %), boreal deciduous broadleaf forest (16.7 %), temperate deciduous broadleaf forest (16.1 %), deciduous needleleaf forest (6.6 %), boreal evergreen needleleaf forest (4.4 %), mixed leaf forest (4.3 %), tropical evergreen broadleaf forest (0.9 %), and tropical deciduous broadleaf forest (0.05 %). Although temperate evergreen needleleaf forests and boreal deciduous broadleaf forests were the two largest components by area prior to 2000, their proportional contributions to the total forest area subsequently declined from 31.3 % to 23.4 % and from 20.0 % to 16.7 %, respectively. Conversely, the proportional representation of temperate evergreen broadleaf and temperate deciduous broadleaf forests expanded, increasing from 18.5 % to 27.5 % and from 9.5 % to 16.1 %, respectively (Fig. 4).

Spatially, the primary forest regions are concentrated in Northeast, Southeast, and Southwest China, whereas forest cover is relatively sparse in Northwest, Central, and East China. Furthermore, evergreen needleleaf and evergreen broadleaf forests are predominantly distributed across southern China. Deciduous needleleaf forests are concentrated in the Greater Khingan Range in the northernmost part of Northeast China, while deciduous broadleaf forests are located mainly in Northeast China and the Qinling Mountains of Central China.

With respect to its temporal evolution, our reconstructed forest dataset faithfully reproduces the long-term dynamics of forest cover in China (Fig. 5). According to statistics from the NFI, China's forest cover has a mean annual growth rate of 1.75 %. The reconstructed forest cover dataset reveals a substantial increase in China's total forest area from 115.28 million ha (Mha) in 1981 to 195.45 Mha in 2023, with a peak of 200.05 Mha observed in 2018 (Fig. 5a). This represents an annualised growth rate of 1.50 %, demonstrating strong agreement with this national benchmark. Furthermore, our dataset accurately captures the distinct historical trajectories of two principal forest categories – broadleaf and needleleaf forests – since the 1980s (Fig. 5b and c). This net increase was primarily propelled by the expansion of temperate and boreal forests (Fig. 4). For example, the area of temperate evergreen broadleaf forest more than doubled between 1981 and 2023, increasing from 17.10 to 46.85 Mha (Fig. 4k). During this period, significant areal gains were also recorded for temperate deciduous broadleaf, temperate evergreen needleleaf, and boreal deciduous broadleaf forests. In contrast, absolute changes in the extent of tropical PFTs, boreal evergreen needleleaf forests, and mixed leaf forests were minimal over the same interval (Fig. 4k).

Between 1981 and 2023, the total area experiencing forest change (gross change) was 166.91 million ha (Mha), which is equivalent to approximately 17 % of China's terrestrial surface. This comprised 132.43 Mha of forest gain (14 % of the national land area) and 58.69 Mha of forest loss (6 % of the national land area). Forests that remained stable, persisting from 1981 to 2023, covered 62.29 Mha. This stable area represents 54 % of the total forest extent in 1981, implying that the remaining 46 % of the original 1981 forest cover underwent some form of change during the study period (Fig. 6q). We observed a prevalent pattern of forest change across several regions of China, characterised by a progression from more accessible areas (i.e., lower elevations near roads) to more remote locations (i.e., higher elevations far from roads). This dynamic often manifested as a core-to-edge expansion of existing forest patches (Fig. 6a–p). We detected severe forest loss on the eastern Qinghai-Tibet Plateau from 1990–1996, which appeared to be driven primarily by commercial logging (Chen et al., 2013). Following the implementation of large-scale ecological restoration programmes initiated in 1998 – specifically, the Natural Forest Protection Programme and the Grain for Green Programme – forest cover exhibited a recovery trend from 2000 to 2008 (Liu et al., 2008). Temporally, both forest loss and gain were continuous dynamic processes throughout the entire period. A prominent peak in forest turnover occurred between 1991 and 1996, culminating in 1994 when the combined area of gain and loss surpassed 24.63 Mha. After 2001, both gain and loss areas exhibited a general downwards trend, albeit with notable fluctuations, and moderate recessions in turnover were observed in approximately 2006, 2010, and 2016. Furthermore, forest loss and gain events tended to occur relatively concurrently (Fig. 6q).

Our analysis revealed that forest loss events are typically short in duration (Fig. S14). More than 30 % of all observed losses persisted for only 5–9 years. In contrast, forest gain is characterised by substantially longer persistence, with a modal duration of 25–29 years. This highlights the long-term stability of large tracts of newly established forests. The statistical distributions of persistence durations for forest gain and loss are markedly different. Loss events are predominantly concentrated in the shorter-duration intervals, whereas periods of gain are more concentrated in the medium- to long-duration brackets. This divergence indicates that newly established forests tend towards greater stability and longevity, whereas forest loss manifests as a more fragmented and ephemeral phenomenon.

We assessed the impact of different PFT forcing datasets on ecosystem simulations by comparing outputs from the LPJ–GUESS model driven with our reconstructed PFT product (hereafter PFTlocal) versus the global PFT map (PFTglobal) from the European Space Agency (hereafter PFTESA) and the MCD12Q1 product (hereafter PFTMCD12Q1). The analysis, exemplified using data for the year 2010, quantifies the resulting differences in key ecosystem variables: GPP, NEE, LAI, and ET (Fig. 7). Here, we focus on the mean differences (PFTlocal minus PFTglobal) during the summer period (June–August) to highlight the primary impacts. The results indicate that the most marked differences in the simulated carbon and water fluxes spatially coincide with regions where the three products show substantial differences in the fractional coverage of tree PFTs. Regions with a higher tree cover fraction in the PFTlocal dataset than in the PFTESA dataset, particularly in northeastern China (Fig. 7e, red), exhibit correspondingly elevated GPP, LAI, and ET, along with a diminished NEE. Consequently, the resulting differentials are positive for the former variables and negative for NEE. Conversely, where the PFTESA dataset indicates greater tree coverage, such as in parts of southwestern China (Fig. 7e, blue), these relationships are inverted. Similarly, the comparison with PFTMCD12Q1 yields analogous results (Fig. 8).

To assess the realism of the simulations, we benchmarked the model outputs against a suite of remote sensing-based products (FLUXCOM GPP, FLUXCOM NEE, GIMMS LAI4g, and GLEAM ET), with all the datasets aggregated to a common 0.1° resolution (Fig. 9 for summer, June–August). In Fig. 9a and b, we first present a direct comparison between the surface fluxes simulated using our reconstructed PFT map (PFTlocal) and the observed benchmarks. This baseline comparison highlights inherent discrepancies attributable to both structural biases in the LPJ–GUESS model and uncertainties within the remote sensing products themselves. Figure 9c and d then allow the impact of the PFT forcing to be isolated by showing the change in absolute simulation error relative to the observational data. Areas of improvement (demarcated in red) and degradation (in blue) in model performance when PFTlocal is used in place of PFTESA and PFTMCD12Q1 are explicitly mapped.

High simulated GPP values (> 0.2 kgCm-2) are concentrated in the forested regions of northeastern China (e.g., the Greater Khingan and Changbai Mountains), the central Qinling Mountains, and southeastern China. Conversely, low GPP values are characteristic of the arid and semiarid regions of the northwest and the Tibetan Plateau (Fig. 9a), where productivity is constrained by water availability and low temperatures. On a macroscale, the simulation accurately captures this geographical distribution. PFTlocal demonstrates a distinct advantage in terms of PFT forcing. Compared with PFTESA, it reduces simulation errors across 77.7 % of the domain (Fig. 9c), with improvements primarily concentrated in Northeast China, North China, and the Qinling Mountains (red areas); conversely, areas of increased error (16.2 %) are sporadically distributed across southern China. The improvement is even more pronounced relative to PFTMCD12Q1, for which error reductions cover 85.2 % of the terrestrial area, whereas increased errors are confined to only 8.4 % of the domain, specifically in Southwest China and Taiwan (Fig. 9d). With respect to the NEE simulations, PFTlocal significantly outperforms PFTESA, with fewer errors across 63.1 % of the domain (Fig. 9c). The most pronounced improvements are observed in the Qinling Mountains (indicated by deep red), whereas areas of increased error (28.1 %, blue regions) are primarily concentrated in Northeast and Southwest China. Relative to PFTMCD12Q1, the extent of improvement expands to 69.7 %, whereas the increase in error (21.2 %) is distributed mainly across the Northeast China, Southwest China, and Taiwan regions (Fig. 9d).

The simulated LAI exhibits a spatial pattern analogous to that of the GPP (Fig. 9a). Relative to PFTESA, PFTlocal reduces errors across 66.9 % of the domain (marked in red in Fig. 9c), with the most pronounced improvements observed in the Northeast China and Qinling regions; conversely, areas of increased error (27.4 %, blue) are primarily situated in southern China. For PFTMCD12Q1, the extent of error reduction expands to 77.3 %, while increased errors (16.8 %) are confined largely to Southwest China and Taiwan (Fig. 9d). The spatial pattern of the simulated ET is consistent with that of the other variables analysed (Fig. 9a). With respect to model improvement (Fig. 9c), PFTlocal reduces simulation errors across 78.7 % of China relative to PFTESA (indicated by extensive red areas), except for certain parts of southern China. Compared with PFTMCD12Q1, this proportion increases to 85.3 % (Fig. 9d). These results indicate that the reconstructed dataset demonstrates distinct superiority over the ESA and MCD12Q1 datasets in simulating major carbon and water fluxes across mainland China.