Improving forest aboveground carbon storage estimation by fusing spaceborne LiDAR features and three-dimensional enhanced spectral indices Authors: Yi Long, Hua Sun, Binbin Wang, Fugen Jiang Published: 2026 ABSTRACT SummaryAccurate forest aboveground carbon (AGC) storage estimation is critical for the global carbon cycle and ecosystem sequestration. Optical data suffer from spectral saturation, and spaceborne LiDAR from spatial discontinuity, both hindering seamless high-precision mapping. This study integrated TECIS LiDAR data with Sentinel-2 imagery to construct three-dimensional enhanced spectral indices (TDESIs) by pixel-wise fusing forest canopy height with optical spectral, textural, and fractional vegetation cover features for machine-learning-based AGC estimation. TDESIs outperformed optical-only models, increasing R2 from 0.666 to 0.750 and reducing RMSE by approximately 14.8%, while alleviating spectral saturation. The best performance was achieved by a random forest model using TDESIs and LiDAR-derived metrics (R2 = 0.764, RMSE = 10.728 Mg C/ha). The resultant AGC map (range 8.03–95.69 Mg C/ha) closely matched field conditions, with enhanced discriminability in high-AGC forest. The TDESIs method enables efficient multi-source data integration for large-scale, high-accuracy AGC monitoring. Introduction Amid escalating global climate change, the carbon cycle and its critical component—forest ecosystems—have drawn increasing attention for their role in climate regulation.1 As the principal terrestrial carbon sink, forests store approximately 45% of global carbon.2 Forest aboveground carbon (AGC) storage is a key indicator of ecosystem carbon sequestration capacity, constituting a dynamic and substantial component of the global carbon cycle.3 Since the Paris Agreement, forest carbon management has become central to more than 90% of countries’ Nationally Determined Contributions.4 Accurate AGC estimation is therefore essential for assessing ecosystem health, quantifying carbon sequestration potential, and supporting carbon-trading frameworks. Traditional field surveys, while accurate, are constrained by limited scalability, high cost, and a lack of dynamic monitoring capability.5,6 Remote sensing offers extensive coverage, cost efficiency, and repeatability, making it indispensable for AGC assessment. Optical systems such as Sentinel-2 and Landsat are widely used for their broad coverage, frequent revisits, and rich spectral information.7,8 However, optical signals primarily capture two-dimensional canopy information and are prone to spectral saturation,9,10,11 especially in closed-canopy forests, where sensitivity to variation in carbon storage declines. Active remote sensing technologies help overcome these constraints. Synthetic aperture radar (SAR) provides all-weather observations and sensitivity to vegetation-structure,12,13 yet its complex scattering mechanisms hinder direct retrieval of three-dimensional parameters. LiDAR systems directly measure vertical structure such as tree height and canopy profiles, offering robust characterization of forest architecture.12,14,15 Compared with UAV-borne and terrestrial platforms, spaceborne LiDAR is particularly suited to large-scale inventory because of its broader coverage.16,17 Its canopy-penetrating ability makes it especially effective in dense forests and provides crucial complementary data for improving AGC prediction accuracy. The launch of China’s Terrestrial Ecosystem Carbon Inventory Satellite (TECIS) in August 2022 represents a major advance in AGC monitoring.18,19 As the satellite mission specifically designed for forest carbon sink observation, TECIS integrates a multi-beam LiDAR with a multi-angle hyperspectral imager, enabling synchronous active-passive measurements.20 Operating in a sun-synchronous orbit with near-global coverage (80°N to 80°S) and a 15-day revisit period (Table 1), it enables consistent monitoring. Compared with earlier missions such as ICESat/GLAS and GEDI, TECIS offers enhanced accuracy and scanning flexibility through its multi-beam design. Its unique synergy of active and passive payloads supports the integrated retrieval of structural and spectral forest metrics. Although extensive studies have employed ICESat/GLAS and GEDI data,21,22 research utilizing TECIS remains limited, and its practical performance in spaceborne LiDAR-based AGC estimation requires further validation.Table 1The key parameters of TECISNameAttributeNameAttributeLaunch dateAugust 4, 2022wavelength1064Nominal altitude506 kmalong-track distance180 mFootprint size25 mground tracks5Pulse repetition40 Hzlaser sampling rate1.2 GHz Table 1The key parameters of TECISNameAttributeNameAttributeLaunch dateAugust 4, 2022wavelength1064Nominal altitude506 kmalong-track distance180 mFootprint size25 mground tracks5Pulse repetition40 Hzlaser sampling rate1.2 GHz The key parameters of TECIS However, spaceborne LiDAR typically acquires discrete, spatially discontinuous footprints, hindering seamless AGC mapping.15,17 This limitation has motivated the complementary fusion of seamless optical data with LiDAR to advance carbon estimation methods. Previous integration strategies include: incorporating LiDAR-derived height into structure-aware vegetation indices23,24; using vertical canopy information to correct optical signals for topographic or structural effects25,26,27; and employing continuous optical imagery to spatially extrapolate discrete LiDAR footprints.28,29 These efforts highlight that fusing structural and spectral information at the feature level is key to improving AGC accuracy. Previous studies indicate that AGC generally arises from the synergy of three ecological dimensions: vertical vegetation structure, horizontal coverage, and physiological status.30,31 While spaceborne LiDAR acquires vertical structure and optical remote sensing captures horizontal coverage and physiological condition,22 each alone provides only a partial representation. To better characterize AGC, this study introduces a framework that integrates forest canopy height (FCH)—derived from spaceborne LiDAR—with optical metrics, including fractional vegetation cover (FVC) and spectral and textural indices. By pixel-wise fusing seamless FCH with optical features, a set of three-dimensional enhanced spectral indices (TDESIs) was constructed for machine-learning-based AGC estimation (Figure 1). The main contributions of this study are as follows.1High-resolution, wall-to-wall continuous surfaces were generated from spaceborne LiDAR footprints. Using Sentinel-2 optical imagery and a machine learning model, seamless raster maps of LiDAR-derived structural parameters were produced, effectively overcoming the spatial-coverage limitations inherent in conventional discrete LiDAR observations.2The TDESIs were developed as a core feature-expansion framework for integrating LiDAR and optical data. By combining vertical structural information with horizontal spectral and textural metrics, these indices produce a higher-dimensional variable set with enriched information content, overcoming the representational constraints of conventional optical features in complex forest canopies.3The utility of TECIS, a Chinese satellite mission, was evaluated and validated for regional-scale AGC estimation, providing a key case study to support the operational use of domestic satellite data in ecological monitoring.Figure 1Detailed workflow for TDESIs 1High-resolution, wall-to-wall continuous surfaces were generated from spaceborne LiDAR footprints. Using Sentinel-2 optical imagery and a machine learning model, seamless raster maps of LiDAR-derived structural parameters were produced, effectively overcoming the spatial-coverage limitations inherent in conventional discrete LiDAR observations. High-resolution, wall-to-wall continuous surfaces were generated from spaceborne LiDAR footprints. Using Sentinel-2 optical imagery and a machine learning model, seamless raster maps of LiDAR-derived structural parameters were produced, effectively overcoming the spatial-coverage limitations inherent in conventional discrete LiDAR observations. 2The TDESIs were developed as a core feature-expansion framework for integrating LiDAR and optical data. By combining vertical structural information with horizontal spectral and textural metrics, these indices produce a higher-dimensional variable set with enriched information content, overcoming the representational constraints of conventional optical features in complex forest canopies. The TDESIs were developed as a core feature-expansion framework for integrating LiDAR and optical data. By combining vertical structural information with horizontal spectral and textural metrics, these indices produce a higher-dimensional variable set with enriched information content, overcoming the representational constraints of conventional optical features in complex forest canopies. 3The utility of TECIS, a Chinese satellite mission, was evaluated and validated for regional-scale AGC estimation, providing a key case study to support the operational use of domestic satellite data in ecological monitoring. The utility of TECIS, a Chinese satellite mission, was evaluated and validated for regional-scale AGC estimation, providing a key case study to support the operational use of domestic satellite data in ecological monitoring. Detailed workflow for TDESIs This approach not only improved estimation accuracy but also enhanced computational efficiency, providing a paradigm for developing high-precision, scalable, and transferable remote sensing methods for AGC estimation. Results Wall-to-wall FCH acquisition results This study was conducted in Xintian County, Hunan Province, China (Figure 2). The peak length served as the direct measure of FCH at each LiDAR footprint (Figure 3A). Ground-based validation (Figure 3B) indicated that FCH retrieval from the spaceborne LiDAR footprints achieved an accuracy of 82.13%, with a root-mean-square error (RMSE) of 1.55 m. Subsequently, Sentinel-2 spectral and textural indices were matched to the LiDAR-derived metrics using the geographic coordinates of LiDAR footprints. A random forest (RF) regression model was employed as a spatial interpolator to generate a seamless, wall-to-wall raster surface of LiDAR-derived structural metrics. The resulting spatially continuous peak length surface provided the initial pixel-level FCH values for the study area (Figure 3C) and was used in all subsequent analyses.Figure 2The study area location, land cover, and sample plot distribution map(A) Overview of Hunan Province and the geographic location of Xintian County.(B) Extent of the Xintian County, showing the distribution of land cover, sample plots, and TECIS footprint points.Figure 3FCH extraction, validation, and mapping based on spaceborne LiDAR footprints(A) Distribution of FCH values derived from TECIS footprint points.(B) Fitting results of FCH against ground validation plots.(C) The derived wall-to-wall continuous spatial distribution of FCH. The study area location, land cover, and sample plot distribution map (A) Overview of Hunan Province and the geographic location of Xintian County. (B) Extent of the Xintian County, showing the distribution of land cover, sample plots, and TECIS footprint points. FCH extraction, validation, and mapping based on spaceborne LiDAR footprints (A) Distribution of FCH values derived from TECIS footprint points. (B) Fitting results of FCH against ground validation plots. (C) The derived wall-to-wall continuous spatial distribution of FCH. Results of AGC calculation and variable extraction Plot-level AGC was summed and expressed per unit area. The resulting AGC ranged from 1.843 to 119.048 Mg C/ha, with a mean of 36.435 Mg C/ha; the standard deviation and coefficient of variation were 19.929 Mg C/ha and 54.69%, respectively (Table 2). A total of 157 spectral and textural indices (SIs), comprising surface reflectance values, vegetation indices, and texture features (Table 3), were derived from Sentinel-2. A seamless set of twenty energy- and geometry-related metrics (Table 4) was obtained from TECIS. To evaluate the effectiveness of TDESIs for AGC estimation, four variable sets were established and compared: (i) SIs; (ii) SIs combined with LiDAR-derived metrics (abbreviated as “SIs + LiDAR-M″); (iii) TDESIs; and (iv) TDESIs combined with LiDAR-derived metrics (abbreviated as “TDESIs + LiDAR-M”).Table 2Plot introduction and AGC calculation for major tree speciesTree species groupPlot numberCompatible modelsAGC (Mg C/ha)Pinus massoniana103AGB=0.052318DBH2.0361H0.66779AGC=0.4648AGB5.391–112.434Cunninghamia lanceolata71AGB=0.065662DBH1.7504H0.78038AGC=0.4786AGB7.287–78.859Quercus14AGB=0.092757(DBH2H)0.89013AGC=0.4280AGB6.257–104.085Other soft broad-leaved trees67AGB=0.058152DBH2.0808H0.56270AGC=(0.4149+0.001312DBH)AGB1.843–119.048Note: AGB denotes forest aboveground biomass (kg), DBH denotes diameter at breast height (cm), H denotes tree height (m), and AGC denotes forest aboveground carbon storage (kg).Table 3Details of spectral and textural indices used in this studyNameFormula/ExpressionSurface reflectanceBi=Bandi,(i = 2,⋯,7,8A,11,12)B2-Blue, B3-Green, B4-Red, B5-RedEdge1, B6-RedEdge2, B7-RedEdge3, B8A- Near infrared (NIR), B11-Shortwave infrared 1 and B12-Shortwave infrared 2Simple ratioSRij = Bandi/Bandj,(i,j = 2,⋯,7,8A;i≠j)Enhanced Vegetation IndexEVI = 2.5(NIR-Red)/(NIR+6.0Red-7.5Blue+1)Normalized Difference Vegetation IndexNDVI=(NIR-Red)/(NIR+Red)Atmospherically Resistant Vegetation IndexARVI = [NIR-(2×Red-Blue)]/[NIR+(2×Red-Blue)]Modified simple ratioMSR=(NIR/Red−1)/NIR/Red+1Soil Adjusted Vegetation IndicesSAVI=[(NIR−Red)(1+0.5)]/(NIR+Red+0.5)Visible Atmospherically Resistant IndexVARI=(Green-Red)/(Green+Red-Blue)Difference Vegetation IndexDVI = NIR-RedRed-Green Vegetation IndexRGVI=(Red-Green)/(Red+Green)Modified Soil Adjusted Vegetation IndexMSAVI=(2NIR+1−(2NIR+1)2−8(NIR−Red))/2Normalized Difference Vegetation Index Red-Edge 1 narrowNDVIre1n=(NIR-RedEdge1)/(NIR+RedEdge1)Normalized Difference Vegetation Index Red-Edge 2 narrowNDVIre2n=(NIR-RedEdge2)/(NIR+RedEdge2)Green Chlorophyll IndexCIgreen = NIR/Green-1Red-Edge Chlorophyll IndexCIRE1 = NIR/RedEdge1-1Inverted Red-Edge Chlorophyll IndexIRECI=(RedEdge3-Red)/(RedEdge1/RedEdge2)Non-Linear Red-Edge IndexNLIre=(NIR2-RedEdge1)/(NIR2+RedEdge1)Mean, variance, entropy, data range, skewnessMeani, Vari,Enti, Dati,Skei,(i = 2,3,4,5,6,7,8A)Mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation based on the gray-level co-occurrence matrixCoMeani,CoVari,Hom,Coni,Disi,CoEnti,Seci,Cori,(i=2,3,4,5,6,7,8A)Table 4Details of energy and geometric features derived from spaceborne LiDAR in this studyNameExpressionNameExpressionWaveform quantile heightHq=∫t0tqp(t)dt∫t0t1p(t)dt,q=5,25,75,95,96,97,98,99interquartile range of relative heightsIQ = H75 - H25Waveform lengthWlen = Peaklast - tfirstpeak lengthPlen = Peaklast - PeakfirstLeading edge lengthWll = Peakfirst - tfirsttrailing edge lengthWtl = tlast - PeaklastTotal Waveform lengthWecho = tlast - tfirsttotal echo energyEecho=∫t0t1Amp(t)dtGround echo energyEground=∫tpeaktendAmp(t)dtcanopy echo energyEcanopy = Eecho - EgroundRelative Ground Echo EnergyRground=EgroundEechorelative canopy echo energyRcanopy=EcanopyEechoHalf-wave Optimal Maximum EnergyHOME = PeakG - Pcentroid–– Table 2Plot introduction and AGC calculation for major tree speciesTree species groupPlot numberCompatible modelsAGC (Mg C/ha)Pinus massoniana103AGB=0.052318DBH2.0361H0.66779AGC=0.4648AGB5.391–112.434Cunninghamia lanceolata71AGB=0.065662DBH1.7504H0.78038AGC=0.4786AGB7.287–78.859Quercus14AGB=0.092757(DBH2H)0.89013AGC=0.4280AGB6.257–104.085Other soft broad-leaved trees67AGB=0.058152DBH2.0808H0.56270AGC=(0.4149+0.001312DBH)AGB1.843–119.048Note: AGB denotes forest aboveground biomass (kg), DBH denotes diameter at breast height (cm), H denotes tree height (m), and AGC denotes forest aboveground carbon storage (kg). Plot introduction and AGC calculation for major tree species Note: AGB denotes forest aboveground biomass (kg), DBH denotes diameter at breast height (cm), H denotes tree height (m), and AGC denotes forest aboveground carbon storage (kg). Table 3Details of spectral and textural indices used in this studyNameFormula/ExpressionSurface reflectanceBi=Bandi,(i = 2,⋯,7,8A,11,12)B2-Blue, B3-Green, B4-Red, B5-RedEdge1, B6-RedEdge2, B7-RedEdge3, B8A- Near infrared (NIR), B11-Shortwave infrared 1 and B12-Shortwave infrared 2Simple ratioSRij = Bandi/Bandj,(i,j = 2,⋯,7,8A;i≠j)Enhanced Vegetation IndexEVI = 2.5(NIR-Red)/(NIR+6.0Red-7.5Blue+1)Normalized Difference Vegetation IndexNDVI=(NIR-Red)/(NIR+Red)Atmospherically Resistant Vegetation IndexARVI = [NIR-(2×Red-Blue)]/[NIR+(2×Red-Blue)]Modified simple ratioMSR=(NIR/Red−1)/NIR/Red+1Soil Adjusted Vegetation IndicesSAVI=[(NIR−Red)(1+0.5)]/(NIR+Red+0.5)Visible Atmospherically Resistant IndexVARI=(Green-Red)/(Green+Red-Blue)Difference Vegetation IndexDVI = NIR-RedRed-Green Vegetation IndexRGVI=(Red-Green)/(Red+Green)Modified Soil Adjusted Vegetation IndexMSAVI=(2NIR+1−(2NIR+1)2−8(NIR−Red))/2Normalized Difference Vegetation Index Red-Edge 1 narrowNDVIre1n=(NIR-RedEdge1)/(NIR+RedEdge1)Normalized Difference Vegetation Index Red-Edge 2 narrowNDVIre2n=(NIR-RedEdge2)/(NIR+RedEdge2)Green Chlorophyll IndexCIgreen = NIR/Green-1Red-Edge Chlorophyll IndexCIRE1 = NIR/RedEdge1-1Inverted Red-Edge Chlorophyll IndexIRECI=(RedEdge3-Red)/(RedEdge1/RedEdge2)Non-Linear Red-Edge IndexNLIre=(NIR2-RedEdge1)/(NIR2+RedEdge1)Mean, variance, entropy, data range, skewnessMeani, Vari,Enti, Dati,Skei,(i = 2,3,4,5,6,7,8A)Mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation based on the gray-level co-occurrence matrixCoMeani,CoVari,Hom,Coni,Disi,CoEnti,Seci,Cori,(i=2,3,4,5,6,7,8A) Details of spectral and textural indices used in this study Table 4Details of energy and geometric features derived from spaceborne LiDAR in this studyNameExpressionNameExpressionWaveform quantile heightHq=∫t0tqp(t)dt∫t0t1p(t)dt,q=5,25,75,95,96,97,98,99interquartile range of relative heightsIQ = H75 - H25Waveform lengthWlen = Peaklast - tfirstpeak lengthPlen = Peaklast - PeakfirstLeading edge lengthWll = Peakfirst - tfirsttrailing edge lengthWtl = tlast - PeaklastTotal Waveform lengthWecho = tlast - tfirsttotal echo energyEecho=∫t0t1Amp(t)dtGround echo energyEground=∫tpeaktendAmp(t)dtcanopy echo energyEcanopy = Eecho - EgroundRelative Ground Echo EnergyRground=EgroundEechorelative canopy echo energyRcanopy=EcanopyEechoHalf-wave Optimal Maximum EnergyHOME = PeakG - Pcentroid–– Details of energy and geometric features derived from spaceborne LiDAR in this study Results of variable screening The relative importance and statistical significance of each variable in relation to AGC were evaluated using the four variable sets. Figure 4 presents the top-ten variables by importance. In the SIs set (Figure 4A), the shortwave infrared band B11 showed the highest importance, with most band-reflectance features ranking highly. In the SIs + LiDAR-M set (Figure 4B), B11 remained dominant, and the four most important variables were all optical, confirming the leading role of spectral metrics in this combination. Within the TDESIs set (Figure 4C), SR45 ranked first, followed by Mean4 and CoMean4. In TDESIs + LiDAR-M (Figure 4D), the top three were SR45, Mean4, and SR42. Overall, after three-dimensional enhancement, features associated with the red and red-edge bands attained higher importance.Figure 4Top ten variables ranked by importance for the four variable setsTop ten variables ranked by importance for the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. ∗p < 0.05 and ∗∗p < 0.01 by permutation test. Top ten variables ranked by importance for the four variable sets Top ten variables ranked by importance for the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. ∗p < 0.05 and ∗∗p < 0.01 by permutation test. Stepwise random forest (SRF) selection (Figure 5) showed that, for all sets, RMSE dropped sharply with the inclusion of the first few key variables, reached a minimum, and then rose as more variables were added, indicating that excess features degrade accuracy. The SIs set produced the smallest feature subset, dominated by red-edge and green-band indices. The SIs + LiDAR-M set retained the largest number of variables, highlighting the strong complementarity between optical and LiDAR features and the need for more predictors to establish accurate relationships. The TDESIs set kept 18 features, mainly near-infrared and red-edge-based indices, achieving better initial accuracy with fewer features than simple stacking. The TDESIs + LiDAR-M set also retained 18 features, largely overlapping with TDESIs, while adding the key LiDAR metric Wtl. This confirms that TDESIs encode the core structural-spectral relationships, while the original LiDAR variables provide supplementary three-dimensional detail.Figure 5Variable screening results for the four variable setsVariable screening results for the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) sets. The dashed line corresponds to the number of variables at which RMSE reaches its minimum; the adjacent text shows the corresponding variable count and the names of the selected variables. Variable screening results for the four variable sets Variable screening results for the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) sets. The dashed line corresponds to the number of variables at which RMSE reaches its minimum; the adjacent text shows the corresponding variable count and the names of the selected variables. Comparison of estimation results Four prediction models—support vector regression (SVR), k-nearest neighbors (kNN), extreme gradient boosting (XGBoost), and RF—were constructed using the four variable sets, with the final hyperparameter settings shown in Figure 6, and their performance was evaluated and compared (Figure 7). In this study, the feature set exerted a substantially greater influence on model accuracy than the modeling algorithm. Across all algorithms, models using TDESIs and TDESIs + LiDAR-M consistently outperformed those using SIs and SIs + LiDAR-M. Although SIs + LiDAR-M surpassed SIs alone, both were inferior to the TDESIs-based sets. RF showed the greatest robustness overall. When TDESIs were used instead of SIs, the RF model improved notably, with R2 increasing from 0.666 to 0.750 and RMSE decreasing from 13.252 to 11.294 Mg C/ha. In contrast, mean percentage error (MPE) varied considerably across different feature-model combinations, with no clear monotonic improvement trend; RF models generally yielded relatively high MPE values. The best overall performance was achieved by the RF model trained on the TDESIs + LiDAR-M set, yielding the highest coefficient of determination (R2) (0.764) and the lowest mean absolute error (MAE) (8.330 Mg C/ha), RMSE (10.728 Mg C/ha), and relative RMSE (rRMSE) (30.19%).Figure 6Hyperparameter settings for models trained with the four variable setsHyperparameter settings for the SVR (A), kNN (B), XGBoost (C), and RF (D) models trained with the four variable sets. The dashed lines indicate the hyperparameter values at which RMSE is minimized; the adjacent text gives the optimal values of the remaining hyperparameters.Figure 7Accuracy evaluation results of different models using the four variable setsAccuracy evaluation results of the SVR, kNN, XGBoost, and RF models when trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. Hyperparameter settings for models trained with the four variable sets Hyperparameter settings for the SVR (A), kNN (B), XGBoost (C), and RF (D) models trained with the four variable sets. The dashed lines indicate the hyperparameter values at which RMSE is minimized; the adjacent text gives the optimal values of the remaining hyperparameters. Accuracy evaluation results of different models using the four variable sets Accuracy evaluation results of the SVR, kNN, XGBoost, and RF models when trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. Scatterplots comparing estimated and observed values across all models reveal systematic underestimation in high AGC areas (>60 Mg C/ha). Models using only SIs (Figures 8A–8D) show a clear downward deviation from the 1:1 line at medium-to-high carbon levels, illustrating spectral saturation. The distribution pattern of the SIs + LiDAR-M models (Figures 8E–8H) remains similar to that of SIs alone. In contrast, models with TDESIs (Figures 8I–8L) substantially reduce high-value underestimation, with points aligning more closely to the ideal line. The TDESIs + LiDAR-M combination (Figures 8M–8P) achieves the optimal and most compact distribution. Notably, the RF model (Figure 8P) aligns closest to the 1:1 line, effectively minimizing underestimation while producing the tightest scatter around the regression line.Figure 8Scatterplots of estimated versus observed AGC for the different models(A–D) Scatterplots for the SVR (A), kNN (B), XGBoost (C), and RF (D) models based on the SIs set.(E–H) Scatterplots for the SVR (E), kNN (F), XGBoost (G), and RF (H) models based on the SIs + LiDAR-M set.(I–L) Scatterplots for the SVR (I), kNN (J), XGBoost (K), and RF (L) models based on the TDESIs set.(M–P) Scatterplots for the SVR (M), kNN (N), XGBoost (O), and RF (P) models based on the TDESIs + LiDAR-M set. Scatterplots of estimated versus observed AGC for the different models (A–D) Scatterplots for the SVR (A), kNN (B), XGBoost (C), and RF (D) models based on the SIs set. (E–H) Scatterplots for the SVR (E), kNN (F), XGBoost (G), and RF (H) models based on the SIs + LiDAR-M set. (I–L) Scatterplots for the SVR (I), kNN (J), XGBoost (K), and RF (L) models based on the TDESIs set. (M–P) Scatterplots for the SVR (M), kNN (N), XGBoost (O), and RF (P) models based on the TDESIs + LiDAR-M set. Predictive performance was evaluated in terms of systematic bias and accuracy. The Friedman test confirmed statistically significant differences (p < 0.05) in residual distributions (Figure 9A). RF-based models exhibited mean residuals closest to zero and were negative, whereas most other models showed positive mean residuals. In pairwise comparisons of systematic bias (Figure 9B), the optimal model (d+RF) yielded mainly negative effect sizes relative to the others, confirming its tendency to overestimate. Effect sizes approached zero when d+RF was compared with other RF models, underscoring the algorithm’s consistent bias pattern across variable sets. For absolute residuals (Figure 9C), the Friedman test also indicated significant accuracy differences (p < 0.05). Models using TDESIs displayed lower median errors and less variability than those based on SIs or SIs + LiDAR-M, although all models produced some extreme outliers. Effect sizes for accuracy comparisons (Figure 9D) were predominantly negative when comparing set TDESIs to SIs and set TDESIs + LiDAR-M to TDESIs, indicating that TDESIs integration substantially reduced prediction errors, with LiDAR-M providing modest further improvement. Model accuracy consistently followed the hierarchy: d > c > b > a. Despite a slight systematic overestimation, d+RF achieved the highest overall accuracy, supporting its selection as the optimal model for AGC estimation.Figure 9Descriptive statistics and significance evaluation of model residuals and absolute residuals(A) Descriptive statistics of residuals for each model.(B) Cohen’s d effect size for systematic bias (based on residuals) between model pairs.(C) Descriptive statistics of absolute residuals for each model.(D) Cohen’s d effect size for precision (based on absolute residuals) between model pairs.The boxplots show the median, interquartile range, and outliers; yellow diamonds indicate the mean values, and the red dashed line marks the zero reference. Descriptive statistics and significance evaluation of model residuals and absolute residuals (A) Descriptive statistics of residuals for each model. (B) Cohen’s d effect size for systematic bias (based on residuals) between model pairs. (C) Descriptive statistics of absolute residuals for each model. (D) Cohen’s d effect size for precision (based on absolute residuals) between model pairs. The boxplots show the median, interquartile range, and outliers; yellow diamonds indicate the mean values, and the red dashed line marks the zero reference. Spatial distribution of AGC The spatial distributions of AGC (Figure 10) and the corresponding prediction standard deviations (Figure 11) were generated using the RF models based on the four variable sets. Xintian County exhibited pronounced spatial heterogeneity: The northern region is predominantly mountainous and has experienced historical wildfires, whereas the southern portion is dominated by agricultural land interspersed with fragmented forest patches. All models consistently showed higher AGC in the north and lower values in the south, aligning with established vegetation distribution patterns. Each model also detected reduced AGC in fire-affected areas, confirming sensitivity to disturbance. Prediction standard deviations from the SIs and SIs + LiDAR-M sets were generally high; the SIs-based model showed persistently high standard deviations across the entire region, and although adding LiDAR-M reduced them, values still frequently exceeded 16 Mg C/ha. In contrast, standard deviations from the TDESIs and TDESIs + LiDAR-M models were largely confined to the 0–16 Mg C/ha range.Figure 10Spatial distribution of AGC estimated by the RF model using the four variable setsAGC maps produced by the RF model trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets.Figure 11Standard deviation maps of AGC predictions from the RF model using the four variable setsStandard deviation maps of AGC predictions generated by the RF model trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. Spatial distribution of AGC estimated by the RF model using the four variable sets AGC maps produced by the RF model trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. Standard deviation maps of AGC predictions from the RF model using the four variable sets Standard deviation maps of AGC predictions generated by the RF model trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. Statistically (Table 5), the regional mean AGC was similar across models (32.33–33.13 Mg C/ha), with the TDESIs-only model yielding the highest mean. Models using TDESIs and TDESIs + LiDAR-M produced larger standard deviations and coefficients of variation, indicating greater mapped spatial heterogeneity. The RF model constructed with the TDESIs + LiDAR-M set showed the widest value range (8.03–95.69 Mg C/ha), reflecting an improved ability to capture spatial extremes. As the optimal model, its outputs also aligned more closely with the sample data, offering a more realistic representation of field conditions.Table 5Statistical results for the spatial distribution of AGC in different modelsModelRange (Mg C/ha)Mean (Mg C/ha)Standard deviation (Mg C/ha)Coefficient of Variation (%)SIs + RF11.40–74.5632.3310.7433.23SIs + LiDAR-M + RF11.36–75.0532.9710.2331.02TDESIs + RF10.67–84.1533.1313.2239.91TDESIs + LiDAR-M + RF8.03–95.6932.7113.4841.21Note: SIs, spectral and textural indices; RF, random forest; LiDAR-M, LiDAR-derived metrics; TDESIs, three-dimensional enhanced spectral indices. The contents of the “Model” column represent the RF model based on the SIs, SIs + LiDAR-M, TDESIs, and TDESIs + LiDAR-M sets, respectively. Table 5Statistical results for the spatial distribution of AGC in different modelsModelRange (Mg C/ha)Mean (Mg C/ha)Standard deviation (Mg C/ha)Coefficient of Variation (%)SIs + RF11.40–74.5632.3310.7433.23SIs + LiDAR-M + RF11.36–75.0532.9710.2331.02TDESIs + RF10.67–84.1533.1313.2239.91TDESIs + LiDAR-M + RF8.03–95.6932.7113.4841.21Note: SIs, spectral and textural indices; RF, random forest; LiDAR-M, LiDAR-derived metrics; TDESIs, three-dimensional enhanced spectral indices. The contents of the “Model” column represent the RF model based on the SIs, SIs + LiDAR-M, TDESIs, and TDESIs + LiDAR-M sets, respectively. Statistical results for the spatial distribution of AGC in different models Note: SIs, spectral and textural indices; RF, random forest; LiDAR-M, LiDAR-derived metrics; TDESIs, three-dimensional enhanced spectral indices. The contents of the “Model” column represent the RF model based on the SIs, SIs + LiDAR-M, TDESIs, and TDESIs + LiDAR-M sets, respectively. Sensitivity analysis of AGC estimation Spatial maps of AGC estimation differences were obtained by comparing the TDESIs-based model with the three other models (Figure 12). The patchy pattern of positive and negative differences across the study area indicated that discrepancies were localized rather than systematic. The areal extent of potential overestimation-correction zones was larger when comparing TDESIs with SIs or SIs + LiDAR-M than with TDESIs + LiDAR-M. Overall, the TDESIs model yielded slightly higher estimates than the SIs model (μ = 0.80 Mg C/ha, σ = 10.62 Mg C/ha). In contrast, its regional total was very close to that of SIs + LiDAR-M (μ = 0.16 Mg C/ha), despite high pixel-level variability (σ > 10 Mg C/ha). This suggested TDESIs primarily refined pixel-level spatial distribution rather than altered the regional total. The mean difference between TDESIs and TDESIs + LiDAR-M was small (0.40 Mg C/ha) and showed much lower variability (σ = 4.08 Mg C/ha).Figure 12Spatial distribution of AGC estimation differences(A) Differences between RF models based on TDESI and SI.(B) Differences between RF models based on TDESI and SIs + LiDAR-M.(C) Differences between RF models based on TDESI and TDESIs + LiDAR-M.μ and σ represent the mean and standard deviation of the estimation differences, respectively. Spatial distribution of AGC estimation differences (A) Differences between RF models based on TDESI and SI. (B) Differences between RF models based on TDESI and SIs + LiDAR-M. (C) Differences between RF models based on TDESI and TDESIs + LiDAR-M. μ and σ represent the mean and standard deviation of the estimation differences, respectively. Sensitivity thresholds were delineated based on the global mean and standard deviation of the pixel-wise difference maps and applied to generate a sensitivity zonation map (Figure 13). The results showed that the overestimation sensitivity and underestimation sensitivity zones were primarily concentrated in the forested core of the northern study area, corresponding to regions with high vegetation cover and inherently higher AGC levels. Conversely, in the fire-disturbed central-southern areas and in regions with sparse forest cover, the differences between models were generally classified as non-sensitivity.Figure 13Sensitivity zonation map of AGC estimation differences(A) Sensitivity zonation map of AGC estimation differences between RF models based on TDESI and SI.(B) Sensitivity zonation map of AGC estimation differences between RF models based on TDESI and SIs + LiDAR-M.(C) Sensitivity zonation map of AGC estimation differences between RF models based on TDESI and TDESIs + LiDAR-M. Sensitivity zonation map of AGC estimation differences (A) Sensitivity zonation map of AGC estimation differences between RF models based on TDESI and SI. (B) Sensitivity zonation map of AGC estimation differences between RF models based on TDESI and SIs + LiDAR-M. (C) Sensitivity zonation map of AGC estimation differences between RF models based on TDESI and TDESIs + LiDAR-M. Zonal statistics were performed to examine estimation differences across different FVC and FCH levels. A unimodal relationship was observed between the mean difference and FVC (Figure 14A), while the standard deviation increased nearly linearly with FVC (Figure 14B). In the TDESIs vs. SIs comparison, the mean difference peaked at 1.97 Mg C/ha within FVC level III. A similar pattern—negative differences in low-FVC areas and positive differences in mid-to-high FVC areas—characterized the TDESIs vs. SIs + LiDAR-M comparison. Across all comparisons, the standard deviation rose linearly with FVC, indicating greater pixel-level divergence in denser canopies. Sensitivity zones were concentrated in high-FVC areas, with level V alone exceeding the cumulative area of all other levels (Figure 14C). FCH exerted a primarily linear moderating effect. In the comparisons of TDESIs with SIs and SIs + LiDAR-M, the mean difference increased steadily with FCH (Figure 14D), suggesting that TDESIs yielded higher estimates in taller forests; the standard deviation across FCH levels (Figure 14E) remained around 10 Mg C/ha. In contrast, the TDESIs vs. TDESIs + LiDAR-M comparison showed mean differences near zero and a much lower standard deviation (2.23–4.81 Mg C/ha) across all FCH levels. Sensitivity zones were distributed uniformly across FCH levels (Figure 14F), with the highest proportion occurring in level II (8–10 m), which corresponds to the predominant forest height in the study area.Figure 14Statistical analysis of estimation differences stratified by FVC and FCH(A) Mean of estimation differences for each FVC level.(B) Standard deviation of estimation differences for each FVC level.(C) Proportion of underestimation and overestimation sensitivity within each FVC level.(D) Mean of estimation differences for each FCH level.(E) Standard deviation of estimation differences for each FCH level.(F) Proportion of underestimation and overestimation sensitivity within each FCH level.μ and σ represent the mean and standard deviation of the estimation differences, respectively. Statistical analysis of estimation differences stratified by FVC and FCH (A) Mean of estimation differences for each FVC level. (B) Standard deviation of estimation differences for each FVC level. (C) Proportion of underestimation and overestimation sensitivity within each FVC level. (D) Mean of estimation differences for each FCH level. (E) Standard deviation of estimation differences for each FCH level. (F) Proportion of underestimation and overestimation sensitivity within each FCH level. μ and σ represent the mean and standard deviation of the estimation differences, respectively. Variable effects of the estimation models Shapley Additive Explanations (SHAP) analysis was applied post-hoc to the four optimal models to quantify the contribution of each feature. The SHAP-based importance rankings and summaries for the four models are shown in Figure 15, revealing distinct patterns in how contributions were concentrated across variable sets. For the SIs set, the top-ranked feature B11 contributed 29.95%, and the top three features together exceeded 53.41%. In contrast, the SIs + LiDAR-M model showed a more dispersed pattern: The top-ranked feature SR63 contributed only 10.99%, and seven features were required to reach a cumulative contribution of approximately 53%. In the TDESIs set, the primary feature SR45 contributed 25.00%, with the top three features surpassing 53.65%. The TDESIs + LiDAR-M model followed a similar trend, its top three features reaching 44.62%; the highest-ranked feature SR68A alone contributed 18.93%. Overall, the most important variables were predominantly red-edge-based indices (e.g., IRECI, NDVIre1n, SR68A, SR45) and texture metrics (e.g., Mean2, CoEnt6, Mean3), which together formed the most informative sources for explaining the spatial variation in AGC.Figure 15Feature importance and summary plots based on SHAP valuesFeature importance and SHAP summary plots for the RF models trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. Feature importance and summary plots based on SHAP values Feature importance and SHAP summary plots for the RF models trained with the SIs (A), SIs + LiDAR-M (B), TDESIs (C), and TDESIs + LiDAR-M (D) variable sets. As shown in the SHAP dependence plots (Figure 16), the three most important features in each model exhibited complex, threshold-driven nonlinear relationships with their SHAP values. Most features correlated positively with SHAP values only after exceeding specific thresholds, exerting a negative influence below these points. Their response patterns varied: SR63, IRECI, and Mean2 showed a monotonic increase, with SHAP values rising steadily once the threshold was surpassed. In contrast, NDVIre1 and SR45 displayed an S-shaped response, where their influence peaked within an optimal intermediate range and diminished at the extremes. In striking contrast, B11 exhibited a strong negative correlation with SHAP values, demonstrating a clear inverse threshold effect.Figure 16SHAP dependence plots for the top three most important variables(A–C) SHAP dependence plots for the first (A), second (B), and third (C) most important variables in the SIs set.(D–F) SHAP dependence plots for the first (D), second (E), and third (F) most important variables in the SIs + LiDAR-M set.(G–I) SHAP dependence plots for the first (G), second (H), and third (I) most important variables in the TDESIs set.(J–L) SHAP dependence plots for the first (J), second (K), and third (L) most important variables in the TDESIs + LiDAR-M set. SHAP dependence plots for the top three most important variables (A–C) SHAP dependence plots for the first (A), second (B), and third (C) most important variables in the SIs set. (D–F) SHAP dependence plots for the first (D), second (E), and third (F) most important variables in the SIs + LiDAR-M set. (G–I) SHAP dependence plots for the first (G), second (H), and third (I) most important variables in the TDESIs set. (J–L) SHAP dependence plots for the first (J), second (K), and third (L) most important variables in the TDESIs + LiDAR-M set. Scatterplots of the top three SHAP-ranked features against estimated AGC (Figure 17) were used to quantify the spectral saturation points. For the SIs-based model, the saturation points of three features were 61.50, 61.33, and 60.42 Mg C/ha, with an average of 61.08 Mg C/ha. These values increased slightly in the SIs + LiDAR-M model to 61.93, 61.74, and 62.04 Mg C/ha, yielding a mean of 61.90 Mg C/ha. By contrast, the TDESIs-based model yielded substantially higher saturation points: 79.36 Mg C/ha for SR45, 77.69 Mg C/ha for Mean2, and 75.84 Mg C/ha for SR68A, with an average of 77.63 Mg C/ha—roughly 27% higher than that of the SIs-based model. This confirms that TDESIs extend the responsive range to higher AGC levels, alleviating spectral saturation. In the TDESIs + LiDAR-M model, the saturation points further increased to 79.93, 83.46, and 87.19 Mg C/ha, averaging 83.53 Mg C/ha—a 7.6% improvement over TDESIs alone—indicating that supplementary LiDAR features enhance estimation accuracy in high-AGC regions.Figure 17Fitted spectral saturation points of AGC for the four variable sets(A–C) Saturation point fitted for the first (A), second (B), and third (C) most important variables in the SIs set.(D–F) Saturation point fitted for the first (D), second (E), and third (F) most important variables in the SIs + LiDAR-M set.(G–I) Saturation point fitted for the first (G), second (H), and third (I) most important variables in the TDESIs set.(J–L) Saturation point fitted for the first (J), second (K), and third (L) most important variables in the TDESIs + LiDAR-M set. Fitted spectral saturation points of AGC for the four variable sets (A–C) Saturation point fitted for the first (A), second (B), and third (C) most important variables in the SIs set. (D–F) Saturation point fitted for the first (D), second (E), and third (F) most important variables in the SIs + LiDAR-M set. (G–I) Saturation point fitted for the first (G), second (H), and third (I) most important variables in the TDESIs set. (J–L) Saturation point fitted for the first (J), second (K), and third (L) most important variables in the TDESIs + LiDAR-M set. Discussion Methodology and validation of AGC estimation framework Conventional optical remote sensing is constrained in AGC estimation by its insensitivity to vertical vegetation structure and susceptibility to spectral saturation in dense canopies.9,17 Spaceborne LiDAR provides a critical alternative by directly measuring three-dimensional structural parameters. China’s TECIS, a mission specifically designed for integrated active-passive forest carbon sink assessment, operates in a 506-km sun-synchronous orbit and carries a full-waveform LiDAR with 20-m footprint spacing, achieving higher vertical resolution in dense canopies than photon-counting systems such as ICESat-2.18,19 Its design explicitly anticipates synergy with high-resolution multispectral data. Here, TECIS LiDAR data were integrated with Sentinel-2 imagery, validating the satellite’s operational potential for regional ecological monitoring and establishing a methodological foundation for improved AGC estimation. FCH is a critical biophysical parameter that characterizes stand structure and links vegetation structure to carbon storage.13,32 Building on evidence that FCH can enhance forest biomass and carbon stock predictions,33,34 this study introduced a framework for constructing TDESIs to estimate AGC. A prerequisite was the generation of a high-quality, continuous FCH product from discrete LiDAR footprints. To suppress terrain-induced waveform broadening and distortion,35,36 a bisection approximation method was applied for waveform decomposition. Ground-based validation (Figure 3B) confirmed an FCH retrieval accuracy of 82.13% (RMSE = 1.55 m) at the footprint level. For wall-to-wall FCH generation, traditional spatial interpolation methods (nearest neighbor, inverse-distance weighting, and ordinary kriging) were compared with machine learning models (kNN, RF, XGBoost) that incorporated Sentinel-2 data. Extensive research has demonstrated that RF regression generally excels in remote-sensing-based forest parameter estimation. RF typically offers higher accuracy and robustness than conventional methods such as kNN, and achieves comparable or superior performance to advanced algorithms such as XGBoost, with favorable trade-offs in efficiency, tuning simplicity, and interpretability.37,38,39 Preliminary comparisons confirmed that RF significantly outperformed other methods in spatial consistency, detail preservation, and overall accuracy. Consequently, the continuous FCH products generated by RF provide reliable, high-quality input for all subsequent analyses. The continuous FCH product was multiplied pixel-wise with spectral indices, textural metrics, and FVC derived from optical imagery. This fusion vertically integrated structural information with horizontal physiological attributes, producing a feature set (TDESIs) with higher information density and clearer biophysical meaning. Comprehensive validation confirmed the superiority of the TDESIs. Variable selection analysis (Figure 5) revealed that a model using only approximately 18 TDESIs features outperformed a stacked-feature model (SIs + LiDAR-M) comprising 30 features, highlighting the representational efficiency of TDESIs. Compared with the SIs-only model, the optimal TDESIs-based model increased R2 from 0.666 to above 0.750. TDESIs also raised the mean spectral saturation threshold of key predictors from 61.08 to 77.63 Mg C/ha, extending sensitivity into high-AGC forests, which is consistent with evidence that canopy height mitigates spectral saturation.24,34 Furthermore, SHAP-based importance analysis of all variables (Figure 18) indicated that 30 of the top 40 most influential variables originated from the TDESIs set, confirming their central role in capturing essential AGC information. In summary, by validating a domestic satellite dataset, refining LiDAR processing, and proposing the TDESIs framework, this study established a logically coherent and substantially improved remote sensing paradigm for AGC estimation, offering an operationally viable pathway toward large-scale, high-accuracy AGC mapping. Finally, although validated using TECIS data in a specific subtropical forest, the TDESIs framework is conceptually transferable. Its core logic—the multiplicative fusion of vertical structure with horizontal spectral data—is sensor-independent and applicable to other spaceborne LiDAR missions (e.g., GEDI, ICESat-2) that provide reliable canopy height products,2,40,41 as well as to other high-resolution optical imagery (e.g., Landsat, GF series). By leveraging open global datasets, future applications will enable rigorous assessment of their generalizability across diverse forest types and facilitate their integration into large-scale operational carbon monitoring.Figure 18SHAP-based importance analysis of all variables SHAP-based importance analysis of all variables Insights from model estimation differences and sensitivity analysis Spatial analysis of the estimation differences revealed that TDESIs improved accuracy primarily by redistributing carbon stocks at the pixel level. The difference map between TDESIs and SIs + LiDAR-M (Figure 12B) showed a near-zero overall mean difference yet a high standard deviation, indicating that TDESIs recalibrated the spatial estimates rather than shifting the regional total. This was achieved by multiplicatively fusing FCH with spectral features, thereby embedding vertical structure into the spectral domain and creating a feature space with higher information density. Consequently, the model adjusted AGC predictions based on pixel-specific structure-spectrum relationships. Furthermore, the markedly lower variability in the differences between TDESIs and TDESIs + LiDAR-M suggested that stacking additional raw LiDAR variables refined predictions within the feature framework already established by TDESIs, rather than introducing an independent informational dimension. The improvements delivered by TDESIs exhibited distinct spatial specificity. Sensitivity analysis (Figures 13 and 14) showed that over- and underestimation-sensitive pixels were strongly clustered in core forest areas characterized by high AGC and dense vegetation cover. This pattern aligned with the FVC-stratified results, where the standard deviation of estimation differences increased linearly with FVC. Together, these findings confirm that high-FVC, high-carbon regions are challenging areas where spectral saturation coincides with structural complexity, and where conventional methods are most prone to systematic bias.11,42 By synergistically integrating horizontal cover with vertical structure, TDESIs strengthened the model’s ability to differentiate within these complex environments. This is consistent with the established understanding that, in sparse or disturbed landscapes, reduced signal complexity lowers model uncertainty and promotes convergence among methods.43 Thus, the optimization delivered by TDESIs is spatially targeted, providing particular value for improving accuracy in carbon-rich forest cores. Although this study confirmed that TDESIs effectively refined spatial AGC patterns, some limitations remained. The localized estimation discrepancies may also reflect finer-scale ecological drivers, such as soil properties, micro-topography, local climate, and disturbance history,11 which were underrepresented in the current framework and likely contributed to residual uncertainty. Future work could extend the integration of vertical and horizontal information by incorporating multi-source environmental datasets, such as high-resolution soil maps and long-term climate records. Such integration would support a more comprehensive driver set, enhancing estimation accuracy and generalizability across heterogeneous landscapes. Limitations of the study While the TDESIs framework improved AGC estimation, several limitations warrant consideration. First, the multi-source datasets lacked strict temporal synchronization—a common source of uncertainty.44 To avoid radiometric inconsistencies that can arise from multi-temporal mosaicking under frequent cloud cover, a single cloud-free Sentinel-2 image was selected. Key vegetation parameters and TECIS-derived vertical structure vary little during the dry season, so a temporal offset of several months does not compromise data validity.45,46 Nevertheless, in regions with persistent cloud cover, pronounced phenological variation, or temporally dispersed field surveys, a multi-temporal median composite would be a more robust alternative. Future studies should aim for near-synchronous data acquisition or employ advanced compositing or time-series methods to further reduce temporal misalignment.47,48 Second, input data uncertainties, notably spaceborne LiDAR errors and interpolation artifacts over complex or heterogeneous canopies,26,28 are propagated into TDESIs construction and AGC retrieval, potentially reducing model robustness. Enhancing FCH accuracy under such conditions—for example, through co-registration with high-resolution digital elevation models or terrain-adaptive interpolation algorithms—remains a priority for strengthening the estimation foundation. Spectrally, red-edge-based indices achieved high feature importance, suggesting their richer physiological information and superior sensitivity.49,50 Future work could compute FVC using red-edge vegetation indices to refine performance. Ideally, uncertainty should be propagated through each processing step51; however, the lack of per-footprint uncertainty metrics for TECIS and the complexity of nonlinear error propagation prevent a full uncertainty cascade here. Future work should use Bayesian or Monte Carlo methods for error propagation. Third, relying primarily on FCH captures the uppermost vegetation layer, missing vertical carbon heterogeneity, particularly in stratified subtropical forests. Although TDESIs substantially reduced high-value underestimation, a residual bias above 100 Mg C/ha persisted. The complementary gains from adding raw LiDAR-M features suggest that supplementary metrics (e.g., sub-canopy parameters) could further improve precision where resources permit. Resource availability Lead contact Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Fugen Jiang (jiangkriging@csuft.edu.cn). Materials availability This study did not generate new unique reagents, organisms, or other materials. Data and code availability •The Sentinel-2 data are accessible from the Copernicus Data Space Ecosystem (https://dataspace.copernicus.eu). The land cover data could be downloaded at https://github.com/LiuGalaxy/CRLC. The TECIS data and field plot datasets for Xintian County, as well as the derived datasets generated in this study, are available from the lead contact upon reasonable request (data sharing restrictions apply).•The code used for variable screening and variable effects analysis is available at https://doi.org/10.6084/m9.figshare.31035349.•Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. •The Sentinel-2 data are accessible from the Copernicus Data Space Ecosystem (https://dataspace.copernicus.eu). The land cover data could be downloaded at https://github.com/LiuGalaxy/CRLC. The TECIS data and field plot datasets for Xintian County, as well as the derived datasets generated in this study, are available from the lead contact upon reasonable request (data sharing restrictions apply). The Sentinel-2 data are accessible from the Copernicus Data Space Ecosystem (https://dataspace.copernicus.eu). The land cover data could be downloaded at https://github.com/LiuGalaxy/CRLC. The TECIS data and field plot datasets for Xintian County, as well as the derived datasets generated in this study, are available from the lead contact upon reasonable request (data sharing restrictions apply). •The code used for variable screening and variable effects analysis is available at https://doi.org/10.6084/m9.figshare.31035349. The code used for variable screening and variable effects analysis is available at https://doi.org/10.6084/m9.figshare.31035349. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This work was supported by the 10.13039/501100012166National Key Research and Development Program of China (2023YFD2201703), the 10.13039/501100001809National Natural Science Foundation of China (62401582, 32471861), and the Postgraduate Scientific Research Innovation Project of Hunan Province (CX20240711). We also thank the editor and anonymous reviewers for their constructive comments. Author contributions Conceptualization, Y.L. and F.J.; methodology, Y.L. and B.W.; software, Y.L. and B.W.; formal analysis, Y.L.; writing – original draft, Y.L.; writing – review and editing, Y.L., H.S., and F.J.; supervision, H.S. and F.J.; funding acquisition, H.S., and F.J. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataSentinel-2 Surface Reflectance (Level-2A)Copernicus Data Space Ecosystemhttps://dataspace.copernicus.euLand coverCross-resolution national-scale land-cover (CRLC)https://github.com/LiuGalaxy/CRLCThe TECIS Level 1 dataChina Center for Resources Satellite Data and ApplicationThis paper Available from lead contact upon reasonable request (data sharing restrictions apply)AGC Field survey dataXintian County (255 plots; from November 2023 to May 2024)This paper Available from lead contact upon reasonable request (data sharing restrictionsapply)FCH Field survey dataXintian County (148 plots; May to July 2024)This paper Available from lead contact upon reasonable request (data sharing restrictionsapply)Software and algorithmsArcGIS 10.6ESRIhttps://www.esri.com/en-us/homeR 4.3.3CRANhttps://cran.r-project.orgPython 3.6Python Software Foundationhttps://www.python.org/Original code for this paperThis paperhttps://doi.org/10.6084/m9.figshare.31035349 REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataSentinel-2 Surface Reflectance (Level-2A)Copernicus Data Space Ecosystemhttps://dataspace.copernicus.euLand coverCross-resolution national-scale land-cover (CRLC)https://github.com/LiuGalaxy/CRLCThe TECIS Level 1 dataChina Center for Resources Satellite Data and ApplicationThis paper Available from lead contact upon reasonable request (data sharing restrictions apply)AGC Field survey dataXintian County (255 plots; from November 2023 to May 2024)This paper Available from lead contact upon reasonable request (data sharing restrictionsapply)FCH Field survey dataXintian County (148 plots; May to July 2024)This paper Available from lead contact upon reasonable request (data sharing restrictionsapply)Software and algorithmsArcGIS 10.6ESRIhttps://www.esri.com/en-us/homeR 4.3.3CRANhttps://cran.r-project.orgPython 3.6Python Software Foundationhttps://www.python.org/Original code for this paperThis paperhttps://doi.org/10.6084/m9.figshare.31035349 Method details Overall workflow This study integrated TECIS LiDAR data with Sentinel-2 imagery to construct three-dimensional enhanced spectral indices (TDESIs) for machine-learning-based AGC estimation. The implementation workflow is as follows: First, discrete TECIS footprint data and Sentinel-2 optical imagery were acquired and preprocessed. Subsequently, Sentinel-2 spectral-textural indices were linked to the LiDAR-derived metrics via the geographic coordinates of LiDAR footprints. A random forest regression model, trained on two-thirds of the footprint data and validated on the remainder, was employed as a spatial interpolator to model the nonlinear relationships and generate a seamless, wall-to-wall raster surface of LiDAR-derived structural metrics. The resulting spatially continuous peak length surface provided the initial pixel-level FCH values for the study area. A pixel-wise multiplicative fusion method that combined FVC, FCH, and spectral and textural indices (SIs) was developed to generate a suite of TDESIs for machine-learning-based AGC estimation. Study area The study was conducted in Xintian County, Hunan Province, China (112°02′–112°23′ E, 25°40′–26°06′ N), covering a total area of 1,022.4 km2 (Figure 2A). The region features complex mountainous terrain, with elevations ranging from 200 to 1,080 m and descending northwest to southeast. Its mid-subtropical humid monsoon climate provides abundant sunshine, a long frost-free period, and distinct seasonality. A nature reserve covering 3,333 ha within the county harbors rich floristic diversity, including more than 800 tree species from 88 families and 106 genera. Historically, forest coverage reached 55.6%, but following the 2022 wildfires it declined to 46.36% by 2024. The highly fragmented forest distribution, driven by the complex mountainous and hilly terrain, poses a significant challenge for remote-sensing-based AGC inversion. Field plot survey data acquisition and processing A systematic 1 km × 2 km grid was established across the study area, yielding 255 forest plots (Figure 2B). Field surveys were conducted between November 2023 and May 2024. Within each 20 m × 20 m plot, all trees with a diameter at breast height (DBH) greater than 5 cm were precisely georeferenced using real-time kinematic technology. Key parameters, including DBH, tree height (H), and species identity, were then measured and recorded for each tree. Individual-tree aboveground biomass (AGB) and AGC were estimated using species- or group-specific compatible models (Table 2). Plot-level AGC was summed and expressed per unit area. Sentinel-2 multispectral imagery data acquisition and processing Sentinel-2 is the only high-resolution multispectral satellite equipped with three red-edge bands, which are highly sensitive to vegetation chlorophyll content, biomass, and stress status, thereby significantly enhancing the retrieval of key vegetation biophysical parameters.49 Combined with a revisit frequency of up to five days, it is ideal for monitoring vegetation dynamics. The Sentinel-2 Level-2A optical imagery used in this study was acquired on 12 March 2024 and sourced from the Copernicus Data Space Ecosystem (https://dataspace.copernicus.eu). This product was pre-processed by European Space Agency using the Sen2Cor processor, providing atmospherically corrected, orthorectified, and cloud-masked surface reflectance data.17 All spectral bands were uniformly resampled to a 20 m spatial resolution to match the field plot size for subsequent analysis. Vegetation indices, composed of linear or nonlinear combinations of spectral bands, have been widely used for assessing vegetation growth conditions.7,52,53,54 Texture features capture the spatial patterns of surface objects and are valuable for image interpretation and parameter extraction.10 A total of 157 spectral and textural indices, comprising surface reflectance values, vegetation indices, and texture features, were derived for subsequent analysis (Table 3). All geospatial datasets were processed in a consistent WGS84/UTM projection. Spaceborne LiDAR data acquisition and processing The Terrestrial Ecosystem Carbon Inventory Satellite (TECIS), launched on 4 August 2022, is the China’s first mission dedicated to integrated active–passive forest carbon sink monitoring. Operating in a sun-synchronous orbit at an altitude of 506 km and an inclination of 97.4°, the satellite provides high spatial resolution and rapid data acquisition capabilities, enabling the effective capture of three-dimensional structural information from forest canopies. The payload includes a high-precision LiDAR system with a pulse repetition frequency of 40 Hz and a sampling frequency of 1.2 GHz. In-orbit calibration confirms an elevation measurement accuracy better than 0.7 m.19 The mission-critical parameters of TECIS are detailed in Table 1. The TECIS Level 1 data products used in this study span from May 2023 to January 2024 and comprise fully georeferenced waveform observations. Processing was performed in Python 3.6, incorporating batch reading, quality control, and geolocation. A noise threshold based on the mean and standard deviation of background noise was applied to isolate valid echoes.55 Echoes were smoothed using Gaussian filtering to enhance waveform trends.56 Non-forest footprints were filtered out by removing waveforms with chaotic patterns lacking a dominant peak, single-peak returns (indicating non-vegetated surfaces), and returns with initial or final peak heights below 4 m (the local minimum canopy height).57,58 This process retained 8,055 locations for analysis (Figure 2B). Waveform decomposition was conducted using a bisection approximation method,58 from which twenty energy- and geometry-related metrics (Table 4) were extracted per footprint. All geospatial datasets were processed in a consistent WGS84/UTM projection. Wall-to-wall LiDAR-derived metrics acquisition The peak length, defined as the interval between the first and last discernible peaks in the echo components,59,60 reflects the propagation process of the laser pulse through vegetation and ground, and is widely used to estimate FCH. In this study, it served as the direct measure of FCH at each LiDAR footprint (Figure 3A). To validate the accuracy of the LiDAR-derived FCH, field surveys were carried out from May to July 2024 at 148 forested LiDAR footprint locations. At each site, the heights of representative trees were measured using a laser altimeter, and the mean value was calculated as the reference FCH. Subsequently, Sentinel-2 spectral and textural indices were matched to the LiDAR-derived metrics using the geographic coordinates of LiDAR footprints. A random forest regression model, trained on two-thirds of the footprint data and validated on the remainder, was employed as a spatial interpolator to model the nonlinear relationships and generate a seamless, wall-to-wall raster surface of LiDAR-derived structural metrics. The resulting spatially continuous peak length surface provided the initial pixel-level FCH values for the study area (Figure 3C) and was used in all subsequent analyses. Calculation of three-dimensional enhanced spectral indices AGC is shaped by three fundamental ecological dimensions: horizontal cover, vertical structure, and physiological condition. While spectral and textural indices provide two-dimensional information, integrating canopy height with spectral-textural data has been shown to significantly improve biomass or carbon estimates.33,34,61 In this study, FVC and FCH were used to represent horizontal density and vertical extent, respectively, whereas spectral and textural indices reflected physiological status. A pixel-wise multiplicative fusion method combining FVC, FCH, and spectral and textural indices (SIs) was developed to generate a suite of three-dimensional enhanced spectral indices (TDESIs). This approach generated a high-dimensional feature set that simultaneously captures structural and biogeochemical forest attributes, providing subsequent models with information-rich and ecologically interpretable inputs. The workflow is illustrated in Figure 1. The FVC was estimated using the pixel dichotomy model based on the normalized difference vegetation index (NDVI).62 The NDVI was first calculated using Sentinel-2 imagery bands, and the FVC was then computed through a piecewise function according to NDVI statistical distribution:63(Equation 1)NDVI=(NIR−Red)/(NIR+Red)(Equation 2)FVC={0NDVINDVI95%where, NIR and Red represent the spectral reflectance of the near-infrared and red bands, respectively, in Sentinel-2 imagery; NDVI5% and NDVI95% denote the 5th and 95th percentiles of the NDVI distribution, defining the minimum and maximum values for FVC calculation. A pixel-wise multiplicative fusion strategy was employed to integrate FVC, FCH, and SIs into a series of TDESIS.64 This integration synthesizes horizontal vegetation density, vertical structural information, and spectral-textural characteristics indicative of species composition and physiological status. The resulting TDESIS quantitatively represent the effective three-dimensional space occupied by actively photosynthetic vegetation within each pixel.(Equation 3)TDESIs=Apixel×FVC×FCH×SIswhere, TDESIs represents the series of three-dimensional enhanced spectral indices, Apixel denotes the pixel area, FVC is the fractional vegetation cover, FCH is the forest canopy height, and SIs refer to the spectral and textural indices extracted from Sentinel-2. Furthermore, to account for the spatial misalignment between pixel boundaries and plot boundaries, a spectral convolution method was implemented to extract variable values at the plot locations:(Equation 4)Rplots=∑i=1nWi∗Riwhere, Wi denotes the value of the i-th pixel within the plot, and Ri represents the corresponding area proportion. Variable screening method The relative importance and statistical significance of each variable set in predicting AGC were assessed using permutation-based importance testing. Feature selection was performed using a SRF approach,10 which initiated with the most important variable and sequentially added features based on their contribution to error reduction. The optimal feature subset for each variable set was then used to construct the final AGC prediction models. Model construction method Four machine learning algorithms—support vector regression (SVR), k-nearest neighbors (kNN), extreme gradient boosting (XGBoost), and random forest (RF)—were employed to evaluate the performance gains of TDESIs across different modeling frameworks, validating their utility as a transferable feature set. Their theoretical foundations and hyperparameter configurations are described below. SVR projects data into a high-dimensional feature space using kernel functions, providing robustness to outliers.65,66 In this study, a radial basis function kernel was employed,67 which is the most widely used kernel function and typically exhibits superior performance. The hyperparameters included the penalty coefficient (cost: 0.1, 1, 10, 100) and the kernel parameter (gamma: 0.01, 0.1, 1, 10). kNN makes predictions based on spatial similarity, which aligns well with the spatial autocorrelation inherent in remote sensing data.7 Two geometrically interpretable distance metrics (Manhattan and Euclidean) and three kernel-based weighting functions (rectangular, optimal, and Gaussian) were tested. The number of neighbors (kmax) was tuned from 1 to 15 to balance spatial information use and overfitting. XGBoost iteratively constructs weak learners to minimize a predefined loss function, providing high efficiency and accuracy for structured data.10,68 A lower learning rate and moderate subsampling promote model stability and diversity.69 In this study, a configuration with a learning rate (eta) of 0.01, a row subsampling ratio (subsample) of 0.6, and a column subsampling ratio (colsample_bytree) of 0.6 was primarily used. The optimal values of maximum tree depth (max_depth: range 3–6) and number of boosting rounds (nrounds: range 1–1000) were determined through systematic tuning. RF aggregates predictions from multiple bootstrapped trees to improve accuracy and control overfitting.70 The number of variables tried per split (mtry) and the number of trees (ntree) are two important hyperparameters.67 To ensure a thorough search while limiting computation, mtry was tested from 1 to the total variable count, and ntree was examined from 1 to 500. All models were developed in the R environment, utilizing the e1071, kknn, xgboost and randomForest packages, respectively. Hyperparameters were optimized through grid search to minimize the root-mean-square error (RMSE). The final settings are shown in Figure 6. Quantification and statistical analysis Software and implementation The mapping of the study area overview map and various spatial distribution maps was completed on the ArcGIS 10.6 platform. LiDAR data processing and feature generation were performed using Python 3.6. Optical imagery feature generation, feature selection, model training, statistical evaluation, SHAP analysis, and spectral saturation analysis were implemented using R 4.3.3. Variable importance calculation Feature selection was performed through an SRF approach,10 which initiated with the most important variable and subsequently incorporated additional feature based on their contribution to error reduction. First, the relative importance of each variable is measured:(Equation 5)%IncMSE=1Ntrees∑t=1Ntrees(MSEtper(j)−MSEtorig(j)MSEtorig(j))×100%where, %IncMSE represents relative importance, Ntrees denotes the total number of trees in the random forest, MSEtorig(j) is the original mean squared error of the t-th tree on the out-of-bag samples, and MSEtper(j) is the mean squared error of the t-th tree on the same out-of-bag samples after permuting the variable. Permutation-based statistical inference was performed to assess the relative importance of predictors. Specifically, for each predictor variable, its values in the out-of-bag samples were randomly permuted multiple times. For each permutation, the relative importance was recalculated, thereby generating a null distribution of relative importance under the condition of no association. The observed relative importance was then compared with this permutation distribution to derive a statistical significance level (p-value) for each variable:(Equation 6)pj=1+∑r=1nrepI(VIjnull(r)≥VIj)nrep+1where, pj is the p value for variable j, nrep is the number of permutation replicates, VIj is the relative importance computed from the original data, VIjnull(r) is the relative importance computed from the r-th permuted dataset, and I( ) is an indicator function that returns 1 if the condition is true and 0 otherwise. Significance levels of variable importance In the final variable importance results, significance levels are indicated by asterisks: ∗ indicates p < 0.05 and ∗∗ indicates p < 0.01 based on the permutation test. Accuracy evaluation and verification The dataset was randomly partitioned, with 70% of the plots allocated to model training and 30% to validation. Model performance was evaluated using the coefficient of determination (R2), mean absolute error (MAE), mean percentage error (MPE), RMSE, and relative RMSE (rRMSE).17(Equation 7)R2=1−∑i=1n(yi−yˆi)2∑i=1n(yi−y¯i)2(Equation 8)MAE=1n∑i=1n|yi−yˆi|(Equation 9)MPE=1n∑i=1n(yi−yˆiyi)×100%(Equation 10)RMSE=∑i=1n(yi−yˆi)2n(Equation 11)rRMSE=RMSEy¯×100%where, yi and yˆi are the observed value and predicted value for the i-th observation, respectively, y¯i is the arithmetic mean of the observed values, and n is the number of samples. Furthermore, a comprehensive statistical comparison was conducted. The Friedman test first determined whether significant differences existed among models, supported by descriptive statistics of residuals and absolute residuals.71 Subsequently, Cohen’s d effect size matrices,72 calculated from both residual types, enabled quantitative pairwise comparisons of systematic bias and accuracy. For systematic bias (assessed by mean residuals), a positive Cohen’s d indicated greater positive bias (underestimation) in Model I relative to Model II; a negative value indicated greater negative bias (overestimation). For accuracy (assessed by absolute residuals), a positive Cohen’s d reflected higher accuracy for Model II than for Model I, whereas a negative value indicated the opposite. Spatial distribution mapping and comparison Spatial AGC distribution maps were generated using the optimal model from each variable set. To examine spatial discrepancies, pixel-level difference maps were produced by subtracting the predictions of the other three models from those of the TDESIs-based model. Pixels with positive differences were classified as potential underestimation-correction zones, and pixels with negative differences as potential overestimation-correction zones. A sensitivity analysis was applied using the global mean (μ) and standard deviation (σ) of the pixel-wise differences. Pixels with differences exceeding μ+1σ were labeled overestimation sensitivity, and those with differences below μ–1σ were labeled underestimation sensitivity. The study area was further stratified by FVC and FCH. FVC was divided into five intervals: I (0–0.2), II (0.2–0.4), III (0.4–0.6), IV (0.6–0.8), and V (0.8–1.0). FCH was grouped into four intervals: I (<8 m), II (8–10 m), III (10–12 m), and IV (>12 m). For each interval, the mean and standard deviation of the model differences were computed, and the areal proportions of the two sensitivity types were quantified and compared. SHAP analysis Shapley Additive Explanations (SHAP), rooted in cooperative game theory, fairly allocates prediction contributions among input variables and provides consistent, model-agnostic local explanations.68,69 SHAP analysis was applied post-hoc to the four optimal models to quantify the contribution of each feature. Mean absolute SHAP values identified the most influential features, and SHAP dependence plots revealed monotonic or nonlinear feature-prediction relationships for key predictors. Spectral saturation point calculation Optical remote sensing is constrained in AGC retrieval by limited sensitivity to vertical structure and susceptibility to spectral saturation in dense canopies. In contrast, LiDAR directly captures three-dimensional structural parameters, helping to mitigate saturation and improve estimation. To quantitatively evaluate how different feature sets respond to spectral saturation, this study derived saturation-point thresholds for key predictors based on semi-variogram theory. For each optimal model, the three features with the highest SHAP importance were selected. Spherical models were fitted to scatterplots of these features against corresponding AGC estimates to characterize their relationship.34(Equation 12)y(x)={C0+C{3x2Sat−x32Sat3},0≤x≤SatC0+C,x>Satwhere, y represents the selected feature, x denotes the predicted AGC, Sat is the spectral saturation point of AGC, C0 is the feature value at x = 0, C is the change rate of the feature with increasing AGC, and C0 + C is the feature value at saturation. By setting b0 = C0, b1 = 3C/2Sat, b2 = -C/2Sat3, x1 = x and x2 = x3, the model transforms into a multiple linear regression whose parameters are estimated via least squares.