The sub-meter optical imagery used in this study was acquired over the study area by GaoFen-2 (GF-2) and GaoFen-7 (GF-7) of the GaoFen satellite constellation. The image collection includes 10 330 GF-2 scenes and 1299 GF-7 scenes acquired during 2023–2024, with summer 2024 imagery used as the primary source and 2023 imagery used to fill remaining spatial gaps. The panchromatic and multispectral (PMS) sensors onboard GF-2 provide a 0.8 m panchromatic band and four 3.2 m multispectral bands, while GF-7, as a stereo-mapping satellite, offers a similar spectral configuration with a 0.8 m panchromatic band and four 3.2 m multispectral bands .

All raw images were processed using a comprehensive workflow, including dehazing and enhancement, geometric and radiometric correction, pan-sharpening, cloud detection, mosaicking, and tiling, to generate a continuous GaoFen image mosaic covering the entire four-province study area. The processed imagery was organized into tiles based on a hierarchical grid system analogous to the Google Maps zoom-level framework. The final processed imagery consists of 13 648 image tiles distributed across Yunnan, Guizhou, Guangxi, and Hunan, with the spatial mosaic shown in Fig. S2 and provincial tile statistics summarized in Table S1.

To characterize the highly heterogeneous patterns of tree distribution across the study area, this study used a tree-crown segmentation dataset derived from sub-meter GaoFen imagery and a U-Net model. The dataset originates from previous tree-mapping efforts . Using the same methodological framework, tree-crown segmentation was extended across the four-province study area, resulting in more than 21 billion mapped trees. Tree crowns with a mapped area of at least 0.64 m2, equivalent to one pixel in the original GaoFen imagery, were retained; this threshold represents the minimum mapping unit of the dataset.

Tree-crown segmentation accuracy was assessed using 919 independent image-interpretation plots, separate from the NFI plots and covering the major forest types and topographic conditions across the study area. Within these plots, 9061 crowns were manually delineated, and pixel-wise comparison of the binary crown masks yielded a precision of 0.832, a recall of 0.853, and an F1 score of 0.843. The R2=0.82 reported by Xu et al. (2026) refers to the plot-level agreement between labeled and predicted tree counts in Guangxi. The tree counts for Southwest China are summarized in Table S1.

Canopy height is a key parameter of forest vertical structure and plays an important role in large-scale AGB estimation . This study used a wall-to-wall canopy-height dataset with a spatial resolution of 0.8 m covering the entire study area. As described by , the dataset was generated using a U-Net model with a ResNet-50 backbone, with sub-meter GaoFen imagery as input and airborne LiDAR-derived CHMs as reference data. The dataset achieved an MAE of 1.94 m and an RMSE of 2.76 m on an independent validation dataset, with no evident saturation in the higher canopy-height range. These errors were substantially lower than those reported by for the canopy-height products of , , and (MAE ≥5.47 m; RMSE ≥6.98 m).

Further accuracy assessment using independent UAV LiDAR data (described in Sect. 2.3.2) yielded an MAE of 1.50 m and an RMSE of 2.82 m for the GaoFen-derived canopy-height dataset, while the other three products showed MAE ≥4.13 m and RMSE ≥5.88 m. The corresponding validation results are shown in Fig. S3.

Ground-truth reference data were derived from China’s 9th National Forest Inventory (NFI) . As an integral component of the national forest resource monitoring system, the NFI provides forest metrics, including forest area, stock volume, and AGB, based on systematically repeated field surveys conducted at fixed sample plots nationwide at five-year intervals. The NFI plots used in this study encompass six provinces – Yunnan, Guizhou, Guangxi, Guangdong, Hunan, and Hubei, which collectively capture the complex topographic gradients and ecological diversity characteristic of Southwest China. NFI plots are typically square in shape with an area of 0.0667 ha (1 mu). Plot-level AGB is estimated from measured stand volume using biomass expansion factors (BEF) specific to tree species and age classes. The temporal coverage of the dataset spans 2014–2018; however, detailed timestamp information for individual plots is unavailable.

The temporal gap of up to one decade between the NFI data and the remote sensing imagery used in this study introduces potential uncertainty . In Southwest China, government-led land-use policies and large-scale ecological restoration initiatives, such as afforestation and reforestation programs, have driven extensive forest regeneration and expansion. Consequently, the region is dominated by young to middle-aged, fast-growing forests, where AGB exhibits rapid and substantial decadal dynamics . Under these conditions, direct use of temporally misaligned datasets for model training or validation may affect the accuracy of AGB estimation.

To identify gross inconsistencies between historical plot AGB and the spectral and structural conditions represented by the recent imagery, we applied an automated cross-source consistency screening workflow. First, an initial data-cleaning step removed plots with incomplete records, missing AGB values, zero or negative AGB, or zero or negative mean tree height. Second, three simple features strongly related to AGB were derived from the high-resolution remote-sensing data for the remaining plots: mean NDVI, crown coverage, and mean canopy height. These features represent the spectral condition, horizontal canopy occupancy, and vertical canopy structure of each plot, respectively.

Third, a parsimonious Random Forest model was constructed using these three variables as predictors and the recorded NFI AGB as the response. To avoid screening plots using in-sample fitted values, five-fold cross-validation was applied. In each iteration, the model was trained using four folds and applied to the remaining fold, ensuring that every plot received an out-of-fold prediction from a model that had not used that plot for training. Plots for which the out-of-fold prediction differed from the recorded NFI AGB by more than 50 % were excluded from subsequent model development. The RF was used only as a preliminary quality-control tool; its parameters and predictions were not transferred to CSMF-Net. Its purpose was to identify pronounced inconsistencies between historical plot AGB and the forest conditions represented by the GaoFen-derived features, rather than to select plots according to CSMF-Net prediction performance or to update the historical NFI AGB values to 2024. This procedure was used as sample-compatibility quality control rather than as a direct temporal correction; it removed plots exhibiting severe structural–label incompatibilities while maintaining the representativeness and diversity of the AGB distribution. However, it could not completely eliminate the uncertainty associated with the temporal mismatch. The final set of selected plots (n=3637) was used for model training (n=2545), validation (n=364), and testing (n=728), and their distribution is shown in Fig. S4.

The evaluation CHM data used in this study were obtained from LIDARNET (http://lidar.pku.edu.cn, last access: 15 October 2026) and covered typical forest scenes in Guangxi Province. The data comprise 18 UAV LiDAR datasets acquired in 2020, covering approximately 1208 ha across Guilin (n=10), Fangchenggang (n=6), and Wuzhou (n=2). The LiDAR point clouds were acquired using Riegl VUX-1, with an average point-cloud density exceeding 10 ptsm-2. The resulting CHMs have a spatial resolution of 0.05 m. Their spatial distribution and a representative example are shown in Fig. S5.

The raw UAV LiDAR data underwent a standardized processing pipeline, including digital terrain modeling, canopy-height extraction, and plot-scale aggregation. We examined all CHM samples and excluded those exhibiting changes unrelated to natural growth (e.g., logging). In total, 3982 CHM validation samples (3591 in forests and 391 outside forests) were obtained and used to evaluate the AGB estimates at the plot level. Only samples with available AGB predictions were included in each product-specific validation, which explains variation in sample size among datasets.

The overall architecture of the proposed Canopy Structure-driven Multi-feature Fusion Network (CSMF-Net) is illustrated in Fig. . It consists of a multi-branch feature extractor, a multi-feature cross-attention module, and a regression head.

First, in the feature extraction stage, the GaoFen optical imagery, tree crown data, and canopy height data for a given plot patch are processed through three parallel encoder branches. A ResNet backbone is employed in each branch to extract deep features from the corresponding data layer. Furthermore, prior to feature extraction on the canopy height data, a gating mechanism is applied. This mechanism performs pixel-wise multiplication of the canopy height data with the tree crown segmentation mask, effectively suppressing background interference and enabling the network to concentrate on the vertical structural information of the canopy.

Subsequently, the extracted features (Fi,Ft,Fh) from the different branches are input into the core Multi-feature Cross-Attention (MCA) module. This module is key to achieving deep fusion of cross-feature information within CSMF-Net. It adaptively learns the interdependencies between spectral and structural features to generate a fused feature representation that is richer in information and more expressive.

Finally, the fused feature is input into a regression head, which consists of multiple fully connected layers. This head outputs the plot-level AGB along with its corresponding heteroscedastic uncertainty.

To achieve effective interaction and informational complementarity between spectral and structural information, we designed a Multi-feature Cross-Attention (MCA) module, the detailed structure of which is illustrated on the right side of Fig. .

The features from the sub-meter GaoFen imagery provide spatial context, while the tree crown and canopy height features explicitly capture the horizontal and vertical structural attributes of the forest. Therefore, within the MCA module, we designated the imagery feature as the Query (Q), leveraging its global contextual information to aggregate relevant information from the other structural features. Simultaneously, we concatenated the canopy height feature with the tree crown feature, followed by separate linear projections to derive the Key (K) and the Value (V).

The cross-attention operation is formally expressed mathematically as follows: 1Q=FiWQ2K=ConcatFh,FtWK3V=ConcatFh,FtWV where WQ, WK, and WV are learnable weight matrices used for feature projection. The attention weights are computed by evaluating the similarity between the Query and the Key, followed by normalization using the Softmax function. These weights are then applied to the Value to generate the fused feature representation. 4Ffused=Attention(Q,K,V)=SoftmaxQKTdkV where dk is the dimension of the Key, used for scaling to stabilize the gradients during training.

To jointly optimize the model’s regression accuracy and its predictive uncertainty, we adopted a composite loss function comprising two components: the R2 loss (LR2) and the heteroscedastic uncertainty loss (Lunc). 5Ltotal=LR2+λLunc where λ is a hyperparameter that balances the contributions of the two loss components. In this study, we set λ to 0.1 based on empirical experiments.

The R2 loss is formulated to directly optimize the model’s goodness-of-fit for AGB by maximizing the R2 between the predicted and true values. It is defined as follows: 6LR2=∑i=1N(yi-yi^)2∑i=1N(yi-y‾)2 where N is the number of samples in a batch, yi and yi^ represent the true and predicted AGB values for the ith sample, respectively, and y‾ is the mean of the true AGB values.

The heteroscedastic uncertainty loss is formulated to capture and explain the discrepancies between the imagery and the plot data, enabling the model to learn the prediction uncertainty associated with the input data. For samples where the model exhibits low prediction confidence, their contribution to the overall loss is proportionally reduced, thereby mitigating the negative impact of noisy or anomalous samples on model training. It is defined as follows: 7Lunc=1N∑i=1Ne-log⁡σi2^⋅yi-yi^2+log⁡σi2^ The first term of this loss function is a Mean Squared Error (MSE) term, which is modulated by an adaptive weight; the second term is a regularization term.

For each pixel, the regression branch outputs the AGB estimate, whereas the uncertainty branch outputs the corresponding input-dependent log variance. The corresponding standard deviation and the lower and upper bounds of an approximately 95 % prediction interval are calculated under an approximately Gaussian error distribution as follows: 8σ^i=exp⁡12log⁡σi2^,9q0.025,i=y^i-1.96σ^i,q0.975,i=y^i+1.96σ^i,10Ui=σ^i=q0.975,i-q0.025,i2×1.96≈q0.975,i-q0.025,i4. In Eq. (), the exponential factor multiplying the squared residual is the inverse of the predicted variance, whereas the additive log-variance term prevents unconstrained variance inflation. Together, these terms estimate the conditional prediction-error variance for each input patch. The square root of this variance defines the 95 % prediction interval and represents the pixel-level predictive uncertainty, allowing uncertainty to vary with the spectral and structural characteristics of the input. The conversion from the 95 % interval width to an equivalent standard deviation follows .

The model was implemented using the PyTorch deep learning framework. During training, we employed the AdamW optimizer , which effectively enhances the model’s generalization capability. The initial learning rate was set to 1×10-4, and a learning rate decay strategy was applied to achieve finer parameter adjustments in the later stages of training. Considering both computational constraints and convergence efficiency, the batch size was set to 32. All experiments were conducted on an NVIDIA RTX 3090 GPU.

The CSMF-Net predictions exhibited strong agreement with the independent NFI observations, achieving R2=0.69, MAE = 18.64 Mg ha⁻¹, RMSE = 28.51 Mg ha⁻¹, and an overall bias of 2.38 % (Fig. a). To further evaluate model performance and uncertainty, we grouped the NFI plots in different classes of AGB ranges and CSMF-Net estimates were compared with multiple established AGB datasets across the classes (Fig. c). Overall, CSMF-Net shows reduced predictive uncertainty, with narrower interquartile ranges than existing datasets. Notably, CSMF-Net demonstrates a better capability in capturing both high and low values. In high-AGB regions (> 200 Mg ha⁻¹), CSMF-Net achieves a median prediction of approximately 190 Mg ha⁻¹, which is 50–60 Mg ha⁻¹ higher than that of conventional products, highlighting its ability to retain high-AGB signals without saturation effects. At 150–200 Mg ha⁻¹ AGB levels, CSMF-Net maintains a median of around 145 Mg ha⁻¹, aligning well with NFI observations, whereas other datasets underestimate field data by 25–30 Mg ha⁻¹. In lower AGB areas (50–100 Mg ha⁻¹), CSMF-Net exhibits a median of roughly 70 Mg ha⁻¹ with an IQR < 25 Mg ha⁻¹, effectively suppressing high-noise fluctuations observed in products such as ESA CCI (IQR of about 90 Mg ha⁻¹). These results indicate that our map provides both accurate and stable AGB predictions across the full AGB range, offering a robust, uncertainty-aware basis for large-scale AGB mapping in the study area. The average changes in median estimates between adjacent NFI AGB intervals were 9.5 Mg ha⁻¹ for Yang Map, 13.3 Mg ha⁻¹ for Yan Map, and 4.5 Mg ha⁻¹ for Cai Map, compared with 32.1 Mg ha⁻¹ for FineKarstAGB. ESA CCI showed a comparatively stronger interval response of 25.1 Mg ha⁻¹, but with wide interquartile ranges across the intervals (85.5–101.5 Mg ha⁻¹).

The input-ablation experiments showed that all three inputs contributed to model performance (Table ). Relative to the corresponding two-input variants, adding GaoFen imagery, tree-crown information, and canopy height reduced RMSE by 9.8 %, 27.0 %, and 41.2 %, respectively, and reduced MAE by 10.7 %, 28.8 %, and 42.7 %, respectively. The full model achieved the highest accuracy (RMSE = 28.51 Mg ha⁻¹, MAE = 18.64 Mg ha⁻¹, and R2=0.69).

The evaluation results for the comparison datasets are presented in Fig. . Existing global- and national-scale AGB products generally exhibited weaker relationships with the NFI observations in the study area, with R2 values ranging from 0.03 to 0.12 and RMSE values ranging from 51.16 to 72.25 Mg ha⁻¹. Specifically, the R2 values were 0.12 for Yang Map, 0.04 for Yan Map, 0.06 for ESA CCI, and 0.03 for Cai Map. The Cai Map also showed a pronounced systematic bias, with an overall bias of 33.66 %. Moreover, the fitted linear relationships had consistently low slopes (0.13–0.32), accompanied by substantial underestimation in high-AGB regions (>200 Mg ha⁻¹), indicating varying degrees of saturation.

To isolate the influence of product year from changes in data availability, the temporal-sensitivity analysis evaluated annual and 2015–2018 median estimates using a fixed product-specific set of NFI plots (Fig. S6). Using the 2015–2018 per-plot median estimates reduced RMSE from 50.95 to 45.90 Mg ha⁻¹ for Yan Map, from 71.09 to 64.54 Mg ha⁻¹ for ESA CCI, and from 73.68 to 66.66 Mg ha⁻¹ for Cai Map; the corresponding MAE reductions were 3.99, 5.77, and 5.52 Mg ha⁻¹, respectively. Temporal mismatch therefore affected the absolute evaluation metrics but did not eliminate the broad differences among the products.

The residual distributions in some higher-slope classes were slightly wider than those at low slopes, but neither their centers nor their spreads increased consistently with slope (Fig. a). Continuous slope was only weakly related to absolute residuals (Pearson r=0.046; Spearman ρ=0.074), indicating a limited slope-related effect on prediction errors. Residual distributions across the eight aspect classes were also generally centered near zero (Fig. b). Mean residuals ranged from -5.27 to 1.30 Mg ha⁻¹ and median residuals from -3.57 to 5.36 Mg ha⁻¹. Although residual dispersion varied somewhat among aspect classes, no consistent directional pattern of overestimation or underestimation was observed.

The National Forest Inventory (NFI) plot data serve as an indispensable source of ground truth for remote sensing–based AGB estimation, yet their application often encounters two major challenges. First, the spatial extent of traditional NFI plots is typically comparable to the pixel size of commonly used medium-resolution remote sensing datasets (e.g., Landsat), which can easily result in incomplete or partial spatial overlap between plots and imagery . This spatial mismatch introduces substantial bias and uncertainty, particularly in regions with complex terrain and highly fragmented landscapes. In this study, by leveraging sub-meter GaoFen imagery, we effectively mitigated this spatial mismatch between field plots and remote sensing data, thereby reducing AGB estimation uncertainty. However, a considerable temporal gap between the NFI data collection and the remote sensing observation period remains a systemic challenge . Despite the implementation of rigorous quality control and filtering procedures, forest ecosystems are inherently dynamic, leading to model uncertainty arising from real AGB changes caused by growth, harvesting, or natural disturbances.

Alongside uncertainties related to plot-data matching, tree-crown segmentation remains subject to uncertainty in dense forests. Although connected crowns are separated through crown-center detection and weighted-distance reassignment , adjacent crowns may still be merged when their centers or visible boundaries cannot be reliably distinguished. The influence of these errors on AGB estimation may be partly reduced because the NFI plots represent diverse canopy conditions and CSMF-Net integrates crown segmentation with spectral and textural information from GaoFen imagery and vertical structural information from canopy height. Nevertheless, these complementary inputs cannot eliminate the underlying crown-segmentation uncertainty. Future work will investigate more advanced crown-deblending methods and three-dimensional canopy information.

To examine the historical stability of the retained plots using information independent of the CSMF-Net inputs, we additionally evaluated the 2235 retained plots that could be matched between NFI8 and NFI9. The annualized logarithmic AGB change between repeated observations was standardized within forest-origin and age groups using robust z-scores. A plot was classified as stable when its AGB change remained within the robust range of its forest group (|z|<2), no Hansen forest loss was recorded during 2014–2024 , no persistent decline was jointly detected by at least two of Landsat NDVI, NBR, and NDMI , and sufficient temporal observations were available. Plots with extreme AGB changes (|z|>3.5) or independent evidence of disturbance were classified as anomalous, while the remaining plots were treated as uncertain. In total, 1852 matched plots (82.9 %) showed stable temporal trajectories. This result indicates that the retained dataset was predominantly composed of plots with consistent historical forest conditions. Nevertheless, it does not demonstrate that every retained plot remained unchanged until the acquisition of the GaoFen imagery, and the screening should be interpreted as consistency-oriented quality control rather than a direct temporal correction of the historical NFI AGB values.

To further quantify the remaining temporal uncertainty, an independent nonlinear growth-slowdown analysis based on repeated NFI8–NFI9 observations estimated median potential AGB increases of 13.45–20.23 Mg ha⁻¹ over assumed intervals of 6–10 years (Fig. S9). These scenario-based estimates were not used to update the NFI labels or FineKarstAGB map, but indicate the possible magnitude of the remaining temporal uncertainty.

In the future, where feasible, temporally synchronized or annually updated NFI datasets should be employed to minimize the influence of this temporal discrepancy on model development and validation.

This study employed GaoFen imagery as the primary data source, whose high spatial and radiometric quality was fundamental to achieving accurate AGB estimation. The raw GF imagery was subjected to geometric and radiometric correction, as well as cloud masking and removal. Nevertheless, a few small areas remained affected by residual cloud contamination; their corresponding AGB estimates were excluded from the final dataset to ensure the reliability of the dataset. Furthermore, the acquisition and processing costs of VHR imagery remain substantial. Although this study achieved seamless coverage across the entire Southwest China region, extending such coverage to larger spatial domains (e.g., national scale) or for long-term temporal monitoring remains challenging. This limitation underscores the trade-off between spatial resolution and geographic or temporal coverage. Future research should therefore explore the integration of alternative high-resolution imagery sources, such as wide-swath GaoFen missions (e.g., GF-1, GF-6) or commercial constellations with high revisit frequency (e.g., PlanetScope), for regional AGB estimation . Expanding both spatial and temporal coverage while maintaining high estimation accuracy will be a key research priority moving forward.

Obtaining spatially and temporally consistent sub-meter imagery over a large region remains challenging. Unlike medium- and coarse-resolution sensors that provide broad-swath and frequently repeated coverage, sub-meter GaoFen imagery requires a large number of individual scenes to achieve wall-to-wall coverage. This difficulty is further increased in the humid and mountainous environment of Southwest China, where frequent clouds, fog, and haze limit the availability of usable observations acquired under comparable conditions. Consequently, the GaoFen mosaic combines scenes acquired on different dates and under varying atmospheric, illumination, viewing-angle, and vegetation conditions. Despite geometric and radiometric corrections, image enhancement, cloud detection, and mosaicking, residual scene- or orbit-related discontinuities remain in some local areas and may propagate into the corresponding AGB estimates. These discontinuities should not be interpreted as abrupt ecological changes, and caution is required when analyzing fine-scale AGB variations across visible image boundaries. Future versions of the dataset will explore multi-temporal radiometric harmonization, improved blending of overlapping observations, and seamline optimization to reduce these artifacts.

This study extended AGB estimation to non-forest landscapes with high spatial heterogeneity, including scattered farmland trees, urban green spaces, and sporadic urban vegetation. By leveraging VHR GaoFen imagery, we achieved fine-scale mapping of AGB in these complex environments. It is important to note that accurate AGB estimation in non-forest scenes relies heavily on VHR data; only such imagery can effectively capture and distinguish these small, heterogeneous vegetation units. However, the validation of AGB estimates in non-forest areas remains a challenge, because NFI plots are mostly found in closed canopy forests. The absence of scale-matched, high-quality ground truth data for validating sub-meter or higher-resolution AGB products remains a critical limitation. Traditional NFI plots, constrained by their coarse spatial scale, cannot adequately represent AGB variation arising from fine-scale structural differences. Future research urgently needs to construct or acquire high-quality, scale-matched validation data specifically for non-forest scenes and VHR AGB products. Future work could involve conducting Terrestrial Laser Scanning (TLS) or Airborne Laser Scanning (ALS) data acquisition, such as LIDARNET, to directly measure AGB across urban trees, agroforestry landscapes, and other structurally complex environments at the sub-meter scale. These TLS and ALS datasets can generate high-density 3D point clouds that enable individual tree-level feature extraction, thereby supporting robust validation of VHR AGB products. Incorporating such high-quality, high-resolution reference data will significantly improve the quantitative accuracy and reliability of AGB estimation in non-forest areas, providing a stronger empirical foundation for urban carbon accounting, agroforestry carbon management, and precision forestry.