The GNPF dataset provides the global distribution of natural and planted forests at 30 m spatial resolution for 2021 (Xiao et al., 2024). Derived from Landsat time-series imagery (1985–2021), the dataset distinguishes forest types based on differences in disturbance frequency, achieving an overall accuracy (OA) of 85 %. The proportions of forest types are consistent with the FAO's Global Forest Resources Assessment 2020, ensuring high thematic reliability. The GNPF dataset was used to generate a forest type mask that separates natural and planted forest pixels. This mask served as a spatial reference for forest age estimation, enabling forest-age characterization for each forest type separately.

A multi-decadal Landsat time series (1985–2024) was constructed by leveraging the extensive data catalog of the GEE platform. We ingested all available Tier 1 Level-2 surface reflectance products from Landsat 4, 5, 7, and 8 (TM, ETM+, and OLI sensors). These products were preprocessed using the LEDAPS and LaSRC algorithms for radiometric calibration and atmospheric correction (Ju et al., 2012; Vermote et al., 2018). To ensure a high-quality temporal stack, we implemented rigorous quality masking – including QA_PIXEL for cloud and shadow removal, alongside aerosol and saturation filters. Furthermore, to maintain radiometric consistency across generations of sensors, Landsat 8 reflectance values were normalized to match the ETM+ sensor using established scaling factors.

Vegetation indices are sensitive indicators of canopy condition and ecosystem successional dynamics. In this study, the Normalized Burn Ratio (NBR) was selected as the core spectral indicator due to its high responsiveness to both abrupt canopy loss and gradual post-disturbance recovery (Bright et al., 2019; Escuin et al., 2008; Ryu et al., 2018). NBR, calculated from the Near-Infrared (NIR) and shortwave infrared (SWIR2) bands, effectively captures structural and moisture changes in vegetation, making it particularly suitable for detecting forest loss and subsequent spectral recovery following disturbances such as fire or logging. The NBR is calculated as follows: 1NBR=NIR-SWIR2NIR+SWIR2.

The CCDC algorithm was executed to decompose the 40-year NBR time series into discrete, chronologically homogeneous segments. This algorithm fits a robust harmonic regression model – incorporating long-term trend and intra-annual seasonal components – via ordinary least squares (OLS) to capture steady-state land surface behavior. Land surface modifications or abrupt disturbances are flagged when a sequence of consecutive observations deviates substantially from the predicted model boundaries beyond a specified statistical threshold (chiSquareProbability = 0.99, Table 2). The mathematical formulation of the CCDC fitting model is expressed as follows: 2ρ^i,x=a0,i+a1,icos⁡2πxT+b1,isin⁡2πxT+a2,icos⁡2πxNT+b2,isin⁡2πxNT Where ρ^i,x represents the predicted value of the ith band of fitted Julian date x; i represents the ith spectral band; T represents the number of days in each year; a0,i represents the intercept coefficient of the ith Landsat band; a1,i and b1,i represent the coefficient of the intra-annual variation term of the reflectance of the ith band; a2,i and b2,i represent coefficients for inter-annual change for the ith band; N represents the number of years.

For each successfully initialized model segment, critical temporal attributes were extracted: the model initialization time (tStart), segment end time, the slope (denoting the spectral recovery or degradation rate), and the RMSE. Crucially, tStart represents the starting date of each fitted harmonic model segment after a detected spectral change. For pixels exhibiting zero structural breakpoints across the entire 1985–2024 archive, the entire timeline was cataloged as a single stable segment, establishing the baseline for assigning older forest age classes (≥40 years). The principal execution parameters for CCDC are detailed in Table 2.

To rigorously evaluate the accuracy of forest age estimates, we established a validation framework integrating high-resolution imagery with dense spectral time-series trajectories (Fig. 2). The global land surface was first partitioned into 57 559 independent 0.5° × 0.5° grid cells to ensure spatial representativeness. Within each grid, a stratified random sampling strategy was applied to select five samples for each forest type, achieving global spatial uniformity and avoiding clustering.

The reference forest age for each sample was determined through a multi-evidence interpretation protocol rather than visual interpretation from a single data source. In this study, forest age was defined as the duration of the current recovery cycle following the most recent detectable major disturbance event. The disturbance timing was primarily constrained using annual Landsat-derived spectral trajectories (NBR and NDVI) and the Global Forest Change (GFC) dataset, while historical high-resolution imagery from Google Earth was used as supplementary evidence to verify canopy conditions and disturbance patterns. In this framework, the GFC loss year provided an independent initial estimate of disturbance timing, which was further evaluated against Landsat spectral breakpoints to identify the onset of the current recovery cycle. To address the limitations of inconsistent image quality and temporal gaps in Google Earth (particularly for the 1985–2000 period), an establishment year was confirmed only when the evidence from spectral trajectories, GFC records, and available spatial imagery reached a consensus. For ambiguous cases – such as pixels with persistent cloud cover, blurry historical snapshots, or unclear natural regeneration signals – the samples were excluded to maintain a high-confidence reference set. Ultimately, we obtained 8085 forest age samples, including 4021 NF samples and 4064 PF samples.

The performance of the Global 30 m Forest Age Map for Natural and Planted Forests (GFAM-N/P) was evaluated using five complementary validation sources, including global interpretation samples and field observations, regional forest age datasets, a global benchmark product, national forest inventory data, and a pixel-level uncertainty assessment. Together, these datasets provide validation across multiple forest biomes, spatial scales, and methodological frameworks.

To evaluate the robustness, global applicability, and parameter sensitivity of our CCDC configuration, we conducted systematic sensitivity analyses across four representative Köppen climate zones using a common parameterization framework to assess whether a globally consistent configuration could provide stable performance under different environmental conditions. This evaluation was extended across various forest developmental stages by categorizing forest age into five representative classes (1–10, 10–20, 20–30, 30–39, and ≥40 years), which were subsequently utilized to derive class-wise PA and UA from the resulting confusion matrices.

First, we assessed the sensitivity of the minObservations parameter (Figs. 10a and 11a–b). The results reveal a consistent nonlinear response pattern across all climate zones. When minObservations increases from 5 to 6, both PA and UA exhibit a pronounced improvement, accompanied by a sharp increase in OA. This indicates that a minimum observational constraint is critical for suppressing noise-induced misclassification in the CCDC segmentation process, particularly in regions with frequent cloud contamination such as tropical forests. However, when the threshold increases beyond 6, the improvement becomes marginal and the accuracy metrics tend to stabilize, suggesting that the model reaches a saturation point where additional observations provide limited incremental information. Based on this behavior, a value of 6 is selected as an optimal global compromise between observational robustness and sensitivity to disturbance events.

Second, the sensitivity to the chiSquareProbability threshold was evaluated (Figs. 10b and 11c–d). At low thresholds (0.7–0.8), both PA and UA remain relatively low and highly variable across age classes, with a pronounced imbalance between omission and commission errors. As the threshold increases to 0.9, classification performance improves moderately but still exhibits residual inconsistencies across climate zones. In contrast, when chiSquareProbability reaches 0.99, all climate zones show a substantial and consistent improvement in both PA and UA, with significantly reduced inter-class variability and improved agreement across age groups. This pattern reflects a transition from an over-segmentation regime, where phenological variability is frequently misclassified as disturbance, to a more conservative detection regime that better preserves estimated forest age structure and reduces systematic misclassification.

Finally, regarding spectral input selection (Fig. 10c), a comparative analysis of various band combinations reveals that the inclusion of the NBR consistently yields the highest OA across all biomes, with gains of 3.4 %–7.8 % over raw reflectance-based models (Source). This demonstrates that NBR effectively enhances the signal-to-noise ratio in identifying stand-replacing disturbances by leveraging the sensitivity of shortwave infrared (SWIR) bands to canopy moisture and structural changes.

Globally, forest age exhibits distinct spatial patterns between NF and PF (Fig. 12) (Wang et al., 2026). NF are broadly distributed across tropical, subtropical, and boreal regions, with older stands concentrated in the Amazon Basin, Central Africa, and northern Eurasia. These areas generally correspond to regions with fewer detectable disturbances during the Landsat observation period. In contrast, PF are predominantly distributed in the mid-latitudes, including East Asia, Europe, Australia, and southeastern South America, reflecting intensive forest management and large-scale reforestation efforts over recent decades (Bennett, 2015). The global map reveals a clear latitudinal gradient: younger forests dominate economically developed and actively managed regions (e.g., East Asia, Europe, and parts of North America), while older stands prevail in equatorial and high-latitude zones. Compared with NF, PF exhibit a markedly younger age structure worldwide. Most PF areas fall within the 1–15-year range, particularly in China, Brazil, and the southeastern United States, where short-rotation plantation systems are common. In contrast, NF contain a substantially higher proportion of mature stands (>30 years), especially in remote tropical and Boreal climate zones dominated by natural succession. These spatial contrasts primarily stem from differences in forest management intensity, reforestation policies, and disturbance regimes. Younger PF ages in subtropical and temperate regions correspond to large-scale plantation rotations, whereas older NF age classes in tropical and boreal areas indicate fewer detectable disturbances during the Landsat observation period. Compared with existing coarse-resolution products such as GAMI (100 m), the 30 m forest age map developed in this study provides a more detailed representation of spatial heterogeneity and better captures fine-scale age gradients across fragmented landscapes.

Building upon the global spatial distribution patterns, we further quantified regional variations in the age structure of NF and PF across six continents – Africa, Asia, Australia, Europe, North America, and South America – to reveal their distinct developmental characteristics and management contexts.

The age structures of NF and PF exhibit pronounced regional contrasts (Fig. 13). For NF, Europe (84.38 %), South America (82.61 %), and North America (80.62 %) show a predominance of the oldest mapped age class (Age 36–40), each typically exceeding 80 %. Africa (66.46 %) and Asia (48.61 %) follow this pattern, whereas Australia exhibits a notable bimodal pattern, with substantial proportions occurring in both the youngest (Age 1–5: 31.36 %) and oldest mapped age classes (Age 36–40: 35.64 %). Except for Australia, young NF stands (Age 1–10) occupy less than 8 % across all other continents. Notably, Asia contains a relatively high proportion of intermediate-age forests (Age 21–25: 21.95 %), indicating a comparatively balanced distribution across different mapped age classes. PF, in contrast, are markedly younger across all continents. In Australia, PF are highly concentrated in the youngest class (Age 1–5: 65.77 %), with only 8.12 % in the oldest (Age 36–40). The dominance of younger age classes in Australian PF is likely associated with intensive plantation management and shorter rotation practices (Strandgard and Béland, 2021). Both Africa and South America exhibit a “bimodal age structure with peaks in young and old classes” structure – Africa (Age 1–5: 31.09 %; Age 36–40: 36.60 %) and South America (Age 1–5: 31.11 %; Age 36–40: 32.59 %). PF in North America (Age 1–5: 17.66 %) and Asia (Age 1–5: 18.58 %) are generally skewed toward younger stands, although the older mapped age classes still account for 34.36 % and 30.09 %, respectively. Europe is the only region where the oldest mapped age class dominates PF (Age 36–40: 53.99 %), while young PF classes remain below 20 %. Regionally, PF in Australia, Asia, and North America are dominated by younger stands, Africa and South America exhibit mixed age compositions, and Europe is characterized by a relatively higher proportion of older mapped plantation age classes.

To evaluate the robustness of the CCDC-based forest age mapping framework, we conducted a systematic sensitivity analysis of two foundational hyperparameters – minObservations and chiSquareProbability – across four major Köppen zones (Arid, Boreal, Temperate, and Tropical). The results indicate that mapping accuracy generally improves with stricter statistical constraints, which enhance stability in breakpoint detection. Among the biomes, the Boreal region consistently yielded the highest overall accuracy, whereas the Arid region exhibited comparatively lower performance, primarily due to its inherently low vegetation cover and high background spectral variability (Almalki et al., 2022). In contrast, the Tropical and Boreal zones demonstrated the highest sensitivity to parameter tuning, implying that dense and structurally heterogeneous forest ecosystems exhibit greater sensitivity to breakpoint detection thresholds (DeVries et al., 2015).

The chiSquareProbability parameter reached the highest overall performance at 0.99 across all zones, indicating that a stringent statistical threshold is essential to suppress spurious breakpoints and minimize commission errors in disturbance detection (Fig. 10). Similarly, increasing minObservations improved model stability below a value of 6, beyond which the performance gains plateaued, identifying a distinct saturation threshold.

Beyond the global optimization of the CCDC baseline hyperparameters, the implementation of the dual-threshold segment-fusion logic directly addresses a long-standing bottleneck in optical time-series forest age estimation: the systematic temporal lag between initial disturbance-related spectral changes and subsequent mathematical model convergence (Tian et al., 2023). Crucially, the tailored spectral recovery slope thresholds of > 0.03 for NF and > 0.04 for PF – were not arbitrarily selected but were empirically derived as the statistical means computed directly from our extensive historical reference sample points. This data-driven differentiation is consistent with the distinct spectral recovery characteristics associated with the two forest origins. For NF, the statistically derived lower mean slope (> 0.03) accommodates the relatively slower and more variable spectral recovery patterns, which are often heavily confounded by heterogeneous canopy conditions and mixed spectral signals that increase trajectory variability (Tian et al., 2015). Conversely, plantation forests often exhibit stronger and more temporally consistent spectral recovery signals following disturbance or harvesting events. By enforcing the more stringent 0.04 threshold for PF, the framework better captures the rapid early-stage recovery characteristics of intensively managed plantations (e.g., Eucalyptus or Pinus) (Deng et al., 2020). This restriction helps avoid premature fusion decisions caused by rapid spectral recovery signals during early plantation development, thereby reducing potential age underestimation. Nevertheless, utilizing a single, globally uniform mean threshold for each forest origin introduces certain geographic limitations. By compressing multi-biome successional dynamics into global average constants, the framework inevitably overlooks intra-type variations driven by macroclimatic gradients. For instance, cold-constrained boreal NF or slow-growing arid plantations may display actual recovery slopes that fall slightly below these statistical means (0.03 or 0.04), potentially leading to localized omissions where recovery segments fail to fuse (Han et al., 2021). Conversely, in hyper-productive wet tropical domains, NF recovery signals might exceed the PF threshold, blurring the algorithmic distinction between NF and PF trajectories. While these generalized thresholds represent an objective trade-off to maintain global comparability and suppress local noise, future studies may further investigate adaptive thresholds that incorporate climatic gradients, forest functional types, and dominant tree species characteristics as complementary approaches for specific regional applications.

Reconciling localized optimization with inter-regional comparability remains a primary challenge in global-scale mapping. Although our sensitivity analysis identified minor regional variations in parameter responses – such as the need for more observations to mitigate cloud interference in tropical domains (Fig. 10a) – we intentionally maintained a uniform parameter framework (Table 2) to ensure spatial consistency and comparability across global forest ecosystems. Although climate-specific parameterization could potentially improve local performance, applying different parameter configurations across climate zones may introduce artificial spatial discontinuities and reduce the consistency of forest age estimates among neighboring regions. As emphasized in seminal large-scale mapping frameworks, employing an internally consistent approach with standardized operational configurations is imperative to filter out geographic definition variations and data input inconsistencies, ensuring that macro-scale characterizations remain robust and comparable across diverse biomes (Hansen et al., 2013). By employing a consistent configuration validated across four distinct biomes, we ensure that the global contrasts in forest age structure revealed in this study primarily reflect regional disturbance histories and management regimes.

The spatial heterogeneity exhibited by the SUI (Fig. 9) is deeply coupled with the interactions between regional forest stand structures, environmental dynamics, and the time-series segment-fitting mechanism of the CCDC algorithm.

The exceptionally low uncertainty (SUI < 5) observed in regions like southeastern China can be attributed to the prominence of managed forests and plantation ecosystems. In these zones, large-scale disturbance or planting events generate distinct, sharp spectral breakpoints, allowing the CCDC algorithm to fit subsequent spectral recovery trajectories with minimal residual noise. Meanwhile, the high-confidence cluster in the southwestern United States and subtropical southern Africa is primarily driven by atmospheric and climatic advantages. These arid and semi-arid regions feature persistently clear-sky conditions and sparse cloud cover, which guarantee a continuous and abundant volume of high-quality Landsat observations. The low seasonal complexity and minimal background noise in these dry ecosystems enable the CCDC time-series inversion to maintain exceptionally tight spectral trajectories, thereby minimizing fitting residuals.

Conversely, the moderate-to-high SUI values (8–13) captured across tropical rainforest domains (e.g., the Amazon Basin and Central Africa) and high-latitude zones (e.g., northern North America and northern Eurasia) are driven by contrasting environmental constraints. Persistent cloud cover in the tropics and high atmospheric moisture content often limit the acquisition frequency of noise-free optical data, inflating the fitting residuals. Furthermore, the multi-layered canopy structures and sub-pixel heterogeneity inherent to natural rainforests mean that partial disturbances often cause subtle spectral variations rather than clean resets, leading to elevated SUI values. Crucially, these localized compounding factors – where extreme data scarcity due to cloud contamination overlaps with intense sub-pixel canopy dynamics – further escalate the fitting error, manifesting as the scattered pixel-level hot spots of maximum uncertainty (SUI > 12, appearing as yellow pixels) observed within the core tropical domains.

In higher latitudes, the maximum mapping uncertainty (SUI > 12) is driven by severe winter conditions. Prolonged snow cover, winter freezing, and intense seasonal phenological variations inject substantial baseline noise into the optical time-series data, inherently increasing the standard deviation of the spectral fitting residuals. This algorithmic behavior is further supported by our independent validation data, where the Boreal climate zone exhibited the highest empirical error (MAE = 3.84 years), whereas the Tropical climate zone showed the lowest error (MAE = 1.15 years), indicating that the constrained fitting uncertainty did not substantially affect the final mapping accuracy.

A rigorous evaluation of our validation results reveals a nuanced structure of estimation uncertainties across distinct forest developmental stages, as explicitly manifested in the joint plot of continuous estimations and probability densities across different observed age groups (Fig. 14). Contrasting with traditional empirical biomass-to-age models that typically suffer from progressive error expansion in mature cohorts, our CCDC-based framework exhibits highly constrained and narrowly binned probability distributions in the upper-age cohort shown in Fig. 14 (30–40 years). This exceptional convergence at the upper age limit directly corresponds to the high UA (93.48 %) reported in our cross-comparison matrix, demonstrating strong consistency in identifying forests without detectable disturbances during the Landsat observation period.

Nevertheless, a minor, vertically scattered trace of data points within this oldest age cohort extends downward toward the baseline (Fig. 14). This trailing underestimation represents a classic algorithmic artifact within long-term time-series segmentation, where anomalous, non-disturbance events – such as extreme localized drought, severe ephemeral frost, or residual cloud-shadow contamination – temporarily distort the spectral trajectory. Even under a stringent statistical constraint (chiSquareProbability = 0.99), these high-magnitude biophysical or radiometric anomalies can occasionally trigger spurious breakpoints, causing the CCDC model to misinterpret an undisturbed mature stand as a newly initiated segment and prematurely reset its estimated age.

Conversely, while the major age cohorts cluster reasonably within their corresponding reference boundaries, the intermediate successional stages – specifically the 10–20 and 20–30 year classes – exhibit relatively higher structural variability and increased downward deviation (Fig. 14). As indicated by the widened interquartile ranges and extended lower tails of the probability density distributions in these middle-aged groups, a proportion of samples show lower estimated ages than the reference observations. The observed downward dispersion in these 10–30 year stands likely reflects the complex spectral responses associated with partial, low-intensity disturbances (e.g., selective logging, surface fires, or insect outbreaks). In young-to-middle-aged natural forests, partial canopy alterations may generate detectable spectral changes without complete stand replacement. Consequently, CCDC may occasionally interpret these spectral variations or subsequent recovery trajectories as new temporal segments, resulting in apparent age resets and increased underestimation uncertainty. These uncertainties should be considered when applying the dataset to analyses involving carbon dynamics and forest structure.

The segment-fusion logic implemented in this study was designed to address the fragmentation of spectrally continuous recovery trajectories during the transition from rapid canopy regrowth to spectral stabilization. In some cases, transient spectral fluctuations during early successional stages may cause CCDC to divide a continuous recovery process into multiple short segments. By merging trajectories with consistent recovery characteristics, the fusion procedure improves the temporal coherence of forest recovery profiles. However, this procedure does not explicitly identify or remove breakpoints induced by non-stand-replacing disturbances. Consequently, partial disturbances remain an important source of uncertainty in the resulting age estimates. In areas affected by selective logging, thinning, insect outbreaks, low-severity fires, or other canopy-altering events, the detected disturbance year may not correspond to complete stand replacement, potentially leading to underestimation of forest age at affected locations. However, the magnitude of this underestimation varies spatially depending on disturbance frequency, intensity, and forest management practices. Therefore, these age biases are not expected to originate from a single parameter configuration but rather from the interaction between disturbance characteristics, spectral recovery trajectories, and the inherent limitations of optical time-series segmentation.

Beyond historical data depth and segment-fitting sensitivities, conceptual and algorithmic lags introduce minor systematic uncertainties. A technical nuance arises from the discrepancy between the tStart and the actual onset of post-disturbance vegetation recovery. Because the GEE implementation of CCDC requires a minimum number of valid observations to initialize a new stable model segment, the estimated tStart inherently reflects the point of model convergence. This introduces a temporal lag whose magnitude depends on regional observation frequency and the availability of valid Landsat observations. Beyond this algorithmic delay, an ecological lag phase exists between disturbance and tree establishment. In many ecosystems, the post-disturbance recovery trajectory includes an initial phase dominated by grasses and shrubs before tree cohorts become spectrally dominant. Therefore, the disturbance-recovery age estimated from Landsat time-series observations may not exactly correspond to the age of dominant tree cohorts. Depending on the timing of spectral breakpoint detection and subsequent vegetation recovery, this temporal discrepancy may result in either slight overestimation or underestimation of tree establishment age. While this ecological lag does not compromise the identified macro-spatial patterns, it should be considered when applying the dataset to analyses involving carbon dynamics, particularly during early successional stages where carbon accumulation rates differ from those of established young forests.

Additional localized uncertainties stem from phenological interference and sensor artifacts. In temperate and boreal biomes, the pronounced seasonality of deciduous species produces spectral signals (e.g., natural annual senescence and rapid spring green-up) that intrinsically overlap with the spectral signatures of disturbance and recovery. While the CCDC algorithm's harmonic regression model is designed to isolate these seasonal cycles, it may still struggle with atypical phenological shifts driven by extreme climatic events, leading to spurious breakpoint detections. Furthermore, in dense deciduous stands, rapid canopy closure during the first few growing seasons may saturate spectral indices like NBR before the stand reaches true spectral stabilization. Finally, radiometric inconsistencies, particularly the scan line corrector (SLC)-off gaps in Landsat 7 imagery, locally distort spectral trajectories and introduce minor biases, notably manifested as image banding effects that disrupt time-series signal stability across certain African regions.

Current global forest age products – such as GFAD, GAMI, and planting year of plantations (PYP) – face inherent limitations in spatial resolution, regional adaptability, and forest-origin-specific age characterization. Most existing datasets operate at coarse spatial resolutions (100 m) and typically do not distinguish forest origin, leading to mixed and often biased age estimates that obscure management-driven differences in forest dynamics. The 30 m GFAM-N/P developed in this study addresses these challenges by offering spatially explicit and origin-specific age estimates for NF and PF. This fine spatial detail enables the detection of subtle regional variations in age structure, particularly across heterogeneous landscapes where natural and planted forests are closely interwoven. Comparative analyses highlight these improvements: our dataset captures fine-scale gradients absent from coarser products, while mitigating data gaps across regions like South America and Europe that are prevalent in biomass-driven or record-limited empirical models (e.g., GAMI and PYP, Fig. 15).

The pronounced spatial heterogeneity and structural contrasts between NF and PF revealed in this study underscore the necessity of shifting from generalized global forestry frameworks to region-specific, precision management strategies. In short-rotation plantation regions (e.g., Australia, parts of Africa, and Southeast Asia) dominated by young PF cohorts (1–15 years), extending rotation cycles and promoting mixed-species configurations could significantly enhance ecological resilience and long-term soil carbon stabilization. Conversely, in regions dominated by mature NF (e.g., Europe and North America), where older forest age classes are prevalent, conservation frameworks should prioritize maintaining natural regeneration, structural diversity, and habitat continuity to support mature forest biodiversity and reduce the risk of carbon losses associated with future disturbances. Asia, characterized by a relatively balanced forest age composition, represents a critical transitional system where protection-oriented and productivity-driven strategies can harmoniously coexist. Ultimately, the high-resolution, origin-specific forest age data provided by GFAM-N/P offer a transparent, reproducible, and robust empirical foundation for optimizing global forest management, improving carbon budget accounting, and informing policy decisions that reconcile climate mitigation with sustainable socio-economic development.