Chongqing, located in the upper reaches of the Yangtze River basin, is one of the economy centers in southwest China (Figure 1). It spans between 105°28’-110°19’E longitude and 28°16’-32°20’N latitude, covering an area of approximately 82,400 km². Renowned as ‘The Mountain City’, Chongqing is defined by its rugged terrain of hills and mountains. The region falls in the humid subtropical zone, experiencing a typical continental monsoon climate. The mean annual temperature ranges from 17 to 18.8°C, with an average annual precipitation of 1000–1400 mm and annual sunshine duration of 1000–1400 hours. The forest types in Chongqing comprise mainly of evergreen broad-leaved forests, secondary and warm coniferous forests, bamboo forests, and evergreen broad-leaved shrubs. The major tree species found in Chongqing include Pinus massoniana, Cunninghamia lanceolata, Cypress, and Quercus spp.

Under-threshold trees (DBH < 5 cm) in 2017 accounted for nearly 23% of all trees in plot data (14697 out of 60786 trees) (Figure 8A), and their estimated DBH distribution for under-threshold trees was consistent with NFI technique protocol, where most values fall around 5 cm (Figure 8B). The distribution mean and median were 4.983 cm and 4.533 cm, respectively, with interquartile ranging from 3.646 cm to 5.687 cm. Some outliers exceeding 10cm may result from measurement errors, such as boundary positioning errors between two adjacent NFI or operation errors in data collection.

The spatial distribution of the 623 NFI plots used in this study was presented in Figure 9. The plots were distributed relatively even across Chongqing to ensure representative spatial coverage. Plot-level AGB values calculated by allometric equations, including under-threshold trees, ranged from 5.15 to 329.81 Mg·ha^-1, with an average of 66.63 Mg·ha^-1. The majority of AGB across these plots fell between 33.38 and 90.48 Mg·ha^-1 (Table 4). The median and mean plot-level AGB, excluding under-threshold trees, were relatively small, with values of 55.08 Mg·ha^-1 and 64.48 Mg·ha^-1. The overall range was from 0.62 Mg·ha^-1 to 329.81 Mg·ha^-1. Plot-level AGB values in western Chongqing were predominantly less than 50 Mg·ha-1, while plot-level AGB values exceeding 150 Mg·ha^-1 were primarily concentrated along a northeast-southwest mountain ranges (Figure 9). When considering under-threshold trees, the increments of plot-level AGB values in more than 500 plots were within 10%, while the plot-level AGB values in the rest plots experienced increments exceeding 10% (Figure 10).

In the RF model, the prediction of AGB primarily relied on spectral bands, with vegetation indices playing a secondary role (Figure 11A). Notably, B12, B4, B3, and B5 exhibited the highest relative importance at 12.1%, 9.5%, 8.9% and 7.7%, respectively, all showing significant negative correlations with AGB. Vegetation indices derived from spectral bands, such as MVI, MSI, and NDII, also demonstrated relatively high importance. In contrast, topographic features including Curvature and Elevation have relatively weaker importance.

The XgBoost model highlighted the critical importance of spectral bands in predicting AGB, with B12, B5, B3, and B4 identified as the most influential features, contributing 19.9%, 9.45%, 8.6%, and 8.5% to the model’s predictive power, respectively (Figure 11B). Among vegetation indices, MVI and MSI also played notable roles, accounting for 5.9% and 5.8% of the feature importance. Topographic indices, including Curvature, Elevation, significantly enhanced AGB prediction. Additionally, biophysical indices such as CWC demonstrated moderate importance, exhibiting positive feedback effects on AGB estimation.

In CatBoost model, spectral band and vegetation indices remain the most significant features (Figure 11C). Among these, B5 (7.1%) and MVI (6.8%) exhibited the highest relative importance, yet they demonstrated opposing feedback effects on AGB. These were followed by TPI in terms of feature contribution, while other topographic indices, such as Curvature and Elevation, showed moderate importance. In contrast, B12, NDII, and B3 were also identified as highly important features in the model.

In the SVM model, B11 (16.6%) and NDVI (15.2%) were the primary contributors, while vegetation indices such as CIgreen (9.5%) and MCARI (9.4%) were of secondary importance. The significance of topographic features remains moderate. Notably, features such as B2 (1.8%), B4 (1.1%), B3 (1.0%), and B5 (0.9%), which exhibited high importance in the other three models, showed relatively low importance in SVM (Figure 11D).

The CV was used to evaluate the stability of the models. Both individual and stacked ensemble model demonstrated robust performance in estimating forest AGB. When including under-threshold trees, the CatBoost model demonstrated the highest predictive accuracy among all individual models, with a mean R² of 0.64 (interquartile range: 0.639–0.647) and a mean RMSE of 25.15 Mg·ha ^-¹ (interquartile range: 24.99–25.28 Mg·ha ^-¹) (Figure 12C). In contrast, XgBoost exhibited the lowest accuracy, with a mean R² of 0.62 (interquartile range: 0.616–0.633) and a mean RMSE of 25.97 Mg·ha ^-¹ (interquartile range: 25.66–26.27 Mg·ha ^-¹) (Figure 12B). The RF model showed intermediate performance, with a mean R² of 0.64 (interquartile range:0.635–0.645) and a mean RMSE of 25.26 Mg·ha ^-¹ (interquartile range: 25.11–25.39 Mg·ha ^-¹) (Figure 12A). SVM’s performance was slightly lower than that of RF, with a mean R² of 0.64 (interquartile range:0.634–0.643) and a mean RMSE of 25.70Mg·ha ^-¹ (interquartile range: 25.48–25.87 Mg·ha ^-¹) (Figure 12D). The stacked ensemble model slightly improved prediction accuracy, with a mean R² of 0.65 (interquartile range: 0.646–0.657) and an RMSE of 24.38 Mg·ha ^-¹ (interquartile range: 24.22 to 24.56 Mg·ha ^-¹) (Figure 12E).

When excluding under-threshold trees, the stacked ensemble model consistently outperformed individual models in prediction accuracy, achieving a mean R² of 0.65 (interquartile range: 0.646–0.656) and a mean RMSE of 25.58 Mg·ha ^-¹ (interquartile range:25.38–25.78 Mg·ha ^-¹) (Figure 13E). Among the individual models, CatBoost achieved a mean R² of 0.64 (interquartile range: 0.632–0.642) and a mean RMSE of 26.41 Mg·ha ^-¹ (interquartile range: 26.23–26.59 Mg·ha ^-¹) (Figure 13C). The RF model exhibited slightly lower accuracy, with a mean R² of 0.63 (interquartile range: 0.629–0.639) and a mean RMSE of 26.53 Mg·ha ^-¹ (interquartile range: 26.32–26.68 Mg·ha ^-¹) (Figure 13A). XgBoost displayed the weakest results, with a mean R² of 0.61 (interquartile range: 0.605–0.623) and a mean RMSE of 27.04 Mg·ha ^-¹ (interquartile range: 26.71–27.36 Mg·ha ^-¹) (Figure 13B). SVM mean R² and RMSE were respectively 0.62 (interquartile range: 0.621–0.632) and 26.80 Mg·ha ^-¹(interquartile range: 26.66–26.93 Mg·ha ^-¹) (Figure 13D).

EV was used to evaluate the final performance of the model. Scatter plots comparing predicted values with plot-level calculated values revealed that all models displayed a systematic bias: they tended to overestimate AGB when values were below 30 Mg·ha ^-¹ and underestimate AGB when it exceeded 125 Mg·ha ^-¹. Among including below-threshold trees, the stacked ensemble model achieved the highest EV accuracy, with an R² of 0.68 and an RMSE of 25.45 Mg·ha ^-¹. Both CatBoost and SVM also performed well, each attaining an R² of 0.66, with RMSE values of 26.33 Mg·ha ^-¹ and 26.86 Mg·ha ^-¹, respectively. RF and XgBoost also demonstrated competitive accuracy (R²: 0.64 and 0.64; RMSE: 26.61 Mg·ha ^-¹ and 26.75 Mg·ha ^-¹) (Figure 14). After excluding under-threshold trees, EV accuracy declined significantly. The R² values of RF, CatBoost, and SVM fell within the range of 0.56–0.57, and all RMSE values exceeded 35.36 Mg·ha ^-¹. The stacked ensemble model somewhat improved predictive performance and reduced discrepancies among the base models (R²=0.59, RMSE=34.52 Mg·ha ^-¹) (Figure 15). It can be concluded that although AGB predictions under both scenarios reached similar stability during model training, the inclusion of under-threshold trees substantially enhances the EV accuracy of the predictions.

The 10m resolution map revealed clear distinct patterns in AGB estimation with and without under-threshold trees. When including under-threshold trees, the average AGB values for the RF, XgBoost, CatBoost, SVM and stacked ensemble models were 61.25 Mg·ha ^-¹,59.15 Mg·ha ^-¹, 60.76 Mg·ha ^-¹, 61.48 Mg·ha ^-¹ and 60.38 Mg·ha ^-¹, respectively, with corresponding total AGB values of 3.58×10⁸ Mg, 3.26×10⁸ Mg, 3.35×10⁸ Mg, 3.39×10⁸ Mg, and 3.33×10⁸ Mg (Figure 16A-E). When excluding under-threshold trees, the average AGB values for RF, XgBoost, CatBoost, SVM and the stacked ensemble model decreased to59.29 Mg·ha ^-¹, 57.08 Mg·ha ^-¹, 58.61 Mg·ha ^-¹, 61.11 Mg·ha ^-¹, and 58.70 Mg·ha ^-¹, respectively, with total AGB values of 3.27×10⁸ Mg, 3.15×10⁸ Mg, 3.23×10⁸ Mg, 3.37×10⁸ Mg, and 3.24×10⁸ Mg (Supplementary Figure S1A-E). The spatial distribution patterns of AGB maps extrapolated from different model implementations (in/ex-clude under-threshold trees data) were similar: lower AGB values were found in the western and northeastern regions of Chongqing, while higher AGB concentrations were mainly located in the southeastern area.

Uncertainty analysis revealed that the uncertainty associated with including under-threshold trees was lower than that when excluding them. When under-threshold trees were included, the average uncertainties for RF, XgBoost, CatBoost, and SVM were 2.68 Mg·ha ^-¹, 5.44 Mg·ha ^-¹, 2.87 Mg·ha ^-¹, and 4.30 Mg·ha ^-¹, respectively. The stacked ensemble model further reduced uncertainty to 3.04 Mg·ha ^-¹ (Figure 17A-E). In the scenario where under-threshold trees were excluded, the average uncertainties for RF, XgBoost, CatBoost, SVM, and the stacked ensemble model were 2.77 Mg·ha ^-¹, 5.47 Mg·ha ^-¹, 2.94 Mg·ha ^-¹, 4.37 Mg·ha ^-¹, and 3.11 Mg·ha ^-¹, respectively (Supplementary Figure S2A-E). Areas with higher uncertainty were primarily distributed in northeastern Chongqing and mountainous regions, while regions with lower uncertainty were mainly concentrated in central and western Chongqing. Additionally, the uncertainties of the RF and CatBoost models were significantly lower than those of XgBoost and SVM, with the stacked ensemble model exhibiting intermediate levels of uncertainty.

According to the technical protocol of the NFI, all trees (including economic trees), bamboo (including bamboo in non-bamboo forests), and tree-like shrubs (excluding shrub-like tree species) in permanent plots should be measured if their DBH reaches 5 cm. Previous studies which utilized NFI data for forest AGB estimation generally did not focus on the impact of trees with DBH less than 5 cm (Zhu et al., 2020; Qian et al., 2021; Li et al., 2022; Zhang et al., 2023), as researchers generally assumed that the biomass contribution of these trees was negligible. However, we compared two consecutive NFI datasets in this study and employed a developed DBH estimation method to estimate the number of trees with DBH less than 5 cm in the earlier measurement. The results indicated that under-threshold trees account for a significant proportion (23%) of the total number of trees measured in 2017, with some plots having AGB variations exceeding 10% (Figure 9 and Figure 10). Although the AGB change of tree plots under threshold value was small, in these plots, the under-threshold trees were primarily broadleaf trees, including Cupressus funebris, Cyclobalanopsis glauca, Pinus massoniana, Cunninghamia lanceolata, and quercus spp. The studies conducted by Yang et al. (2012) and Wen et al. (2015) found that in southern and southwestern China, new wild tree species began to emerge in the forests of Pinus massoniana, Cyclobalanopsis glauca and other broad-leaves forest after the middle-aged forest stage, and community succession existed in over-mature forests to a certain extent. Based on the 2017 forest resource survey data from our study, middle-aged forests, near-mature forests and over-mature forests collectively occupied a significant proportion (42%) of the forest sub-compartments. In addition, excluding under-threshold trees in AGB estimation resulted in noticeably lower R² values and higher RMSE values across all models, indicating a significant decline in prediction accuracy (Figure 14 and Figure 15). These results underscored the non-negligible importance of under-threshold trees in AGB estimation.

In the forest AGB predicted by including under-threshold trees, higher values were primarily concentrated in the Daba, Wulingshan, and Dalou mountain regions. Firstly, these three areas feature high altitudes and inherently possess high forest coverage, leading to elevated forest AGB. Secondly, the 2015 “National Ecological Function Zoning” designated these regions as important areas for water conservation and biodiversity protection, strengthening the protection of existing nature reserves and the management of natural forests. For damaged ecosystems, efforts have been made to adhere to natural restoration, rejuvenate the tree, shrub, and grass vegetation system of evergreen broad-leaved forests, optimize the structure of forest ecosystems, continue implementing the Grain for Green Program and grassland restoration projects, as well as rocky desertification control projects, further enhancing forest coverage. Thirdly, Chongqing has adopted management measures such as establishing multiple nature reserves and forest parks to reduce the impact of human activities, which contributes to increased forest AGB. Areas with lower forest AGB were mainly distributed in the western part of Chongqing, primarily due to frequent human activities and low forest coverage.

Feature variables were pivotal in forest AGB estimation. In the study, spectral bands emerged as the most critical features across all models, with B3, B5, and B12 demonstrating consistently high importance, corroborating findings by Wai et al. (2022) (Figure 11). B5, a red-edge band, exhibited strong capabilities in detecting key vegetation physiological parameters, including chlorophyll content and canopy architecture, which were essential for accurate AGB estimation (Zhang et al., 2023). In contrast, other red-edge bands (B6, B7) have shown greater significance in previous studies (Yang et al., 2012). B3, located at the chlorophyll reflection peak (500–600 nm), was highly sensitive to vegetation “greenness” and effectively reflected physiological states such as chlorophyll content and photosynthetic activity. Similarly, B2 and B4, also within the visible spectrum, were notably important in RF and XgBoost models (Wen et al., 2015). The effectiveness of shortwave infrared (SWIR) bands, particularly B12, in forest AGB estimation had been well-documented. B12 was sensitive to vegetation water content. As biomass and vegetation coverage increased, reflectance absorption by vegetation or water reduced reflectance, whereas low biomass areas exhibited higher reflectance, explaining the negative correlation between B12 and AGB (Wai et al., 2022). These findings underscored the high sensitivity of SWIR, visible, and red-edge bands to biomass, highlighting their critical role in biomass assessment.

The importance of vegetation indices varied across models. MVI was consistently significant across all three models, primarily reflecting forest canopy characteristics that contribute to biomass accumulation. Indices such as MSI and NDII also showed notable importance. MSI typically exhibited a negative feedback effect on AGB, driven by vegetation’s dependence on water and the impact of water stress on growth and physiological processes. NDII and SR primarily indicated changes in foliar chlorophyll and carotenoid content, reflecting photosynthetic activity, becoming key parameters for AGB remote sensing retrieval (Richardson et al., 2002; Merzlyak et al., 2003; Main et al., 2011).

Topographic features, including TPI, Curvature, and Elevation, were strong predictors of forest AGB, particularly in the CatBoost model. Chongqing’s mountainous terrain, characterized by diverse geomorphological features, influences sunlight exposure and water retention, thereby affecting vegetation growth (Wai et al., 2022). TPI and curvature effectively captured these land surface variations. Elevation and aspect also showed significant importance, consistent with previous research (Chen et al., 2019b; Wang et al., 2021). In mountainous regions, variations in elevation and aspect impact moisture levels, temperature, and species richness, ultimately influencing vegetation biomass (Shen et al., 2014; Cong et al., 2019).

Among the individual employed models evaluated, CatBoost demonstrated the best performance, achieving an R² of 0.66, followed by RF and XgBoost, while SVM yielded comparatively lower results (Figure 9). Previous studies have indicated that tree-based machine learning methods were particularly well-suited for ecological remote sensing research (Belgiu and Dragut, 2016). The RF, CatBoost and XgBoost models evaluated in this study were all ensemble methods based on decision trees, and the AGB maps generated by them have a high degree of spatial consistency. Compared with RF and XgBoost, CatBoost was a better choice for estimating AGB due to its advanced design and functions. CatBoost employed an ordered boosting mechanism that reduced the risk of overfitting and minimized the impact of noisy data by processing training examples in a specific order. Given that NFI data often contain inherent noise and variability, CatBoost’s robust handling of categorical features and superior generalization capabilities, makes it particularly well-suited for accurately estimating AGB in complex and noisy datasets. Based on individual models, a stacked ensemble model for AGB estimation using RR as the meta-model was developed. The stacked ensemble model significantly improved prediction accuracy, generalization capability, and robustness, achieving an R² of 0.68 and highlighting its superior performance (Figure 14). The selection of RR as the meta-model was based on the following considerations. Compared to alternative meta-models such as LM, KNN, and entropy weighting, RR exhibited superior and highly consistent performance in both CV and EV, with no evidence of overfitting (Supplementary Table S3). Furthermore, since the base models—RF, XgBoost, and CatBoost—were all tree-based models, their predictions were prone to high correlation (collinearity). RR effectively mitigated collinearity through L2 regularization, yielding more stable and reliable coefficient estimates, although RR’s CV accuracy was slightly lower than that of LM. KNN achieved the highest CV performance, the test performance of KNN declined markedly—indicating overfitting and disqualifying it as a suitable meta-model. The entropy weight method performed similarly to LM and RR on CV, though marginally worse. However, this study found that although the stacked ensemble model demonstrated the highest performance and stronger generalization capability, its uncertainty was greater than that of RF and CatBoost. The primary reason for this is that the uncertainty of the stacked ensemble model is influenced not only by factors such as 5-fold cross-validation but also by performance variations among the base models. Therefore, future research should focus on developing stacked ensemble models that achieve high performance while maintaining low uncertainty. Additionally, the scarcity of field-measured data in high-altitude areas where are usually heavily vegetated resulted in insufficient training, leading to higher uncertainty across all models in these regions. This study also evaluated LM, ANN, and RR as base models and found their performance substantially lower than that of the tree-based ensemble (Supplementary Table S4). These results underscore the strong potential of tree-based machine learning models in AGB estimation.

To validate the generated 10 m spatial resolution forest AGB map, we compared it (including under-threshold trees) with existing AGB products (Figure 18). Our results aligned closely with those of Chang et al. (2021) in six studies (Avitabile et al., 2016; Su et al., 2016; Baccini et al., 2018; Huang et al., 2021; Santoro and Cartus, 2021). The mixed model RMSE in Chang’s study ranged between 24.3 Mg·ha ^-¹ and 29.6 Mg·ha ^-¹, while the stacked ensemble model RMSE in this study was 25.45 Mg·ha ^-¹, indicating comparable model accuracy (Figure 14E). The reason for the differences might be the variations in the base model and the stacking method used. The model exhibited a slight overestimation at low AGB levels (<30 Mg·ha ^-¹) but a significant underestimation at high AGB levels (>125 Mg·ha ^-¹), consistent with documented saturation thresholds in AGB estimation that vary depending on remote sensing data, modeling approaches, and forest structure (Chen et al., 2018; Qian et al., 2021; Wai et al., 2022). First, remote sensing data limitations contributed to these errors. In low AGB areas, the dense canopy structure of small trees obscures thinner trunks (smaller DBH). The spatial resolution of Sentinel-2 (10–20 m) primarily captures spectral characteristics of leaves to estimate AGB, failing to adequately represent trunk structures, which leads to overestimation. Conversely, spectral saturation in high-biomass regions (particularly dense forests) reduced the ability of sensor to discriminate subtle vegetation differences. Although red-edge band of Sentinel-2 partially mitigated saturation, improper band combinations or model selection could still result in underestimation at high AGB levels. Second, model training limitations introduced additional biases. The scarcity of high-biomass samples in the training dataset led to insufficient learning of extreme values, causing the model to regress toward the mean and underestimate high AGB. Meanwhile, low AGB may be overestimated due to noise or mixed-pixel effects (e.g., soil background interference). Therefore, Addressing AGB saturation remained a significant challenge in remote sensing (Qian et al., 2021). Potential improvements to mitigate AGB underestimation include leveraging hyperspectral imagery and LiDAR data to construct three-dimensional forest models, as well as integrating climate and environmental data to enhance biomass estimation accuracy (Feng et al., 2024). Additionally, exploring novel methods such as parametric decomposition and clustering to characterize horizontal and vertical forest structure details could provide alternative approaches.

Computational performance considerations: For the AGB prediction using stacked ensemble models in this study, the computing device must have a minimum of 64 GB RAM and over 500 GB storage memory. The runtime for a single model under a specific scenario exceeds 24 hours. Therefore, it was recommended to implement this method on high-performance computers supporting parallel processing capabilities.