The forest is the most important terrestrial ecosystem on Earth, playing a critical role in the global carbon cycle and terrestrial biosphere (He et al., 2007). The aboveground biomass (AGB) of forest is an important parameter that characterizes their carbon sequestration capacity. Because forest biomass affects a range of ecosystem processes, such as carbon and water cycles, energy fluxes, and thus local and regional climate, the development of sustainable forest management strategies requires accurate information on forest AGB (Urbazaev et al., 2018). Quantitative estimates of aboveground forest biomass provide basic data to support the carbon cycle of the global forest biomass system, thus contributing to the development of global carbon reduction policies and climate change mitigation and providing the necessary information for the development of sustainable forest management (Dixon et al., 1994; Cao et al., 2014). Forest AGB is all aboveground living material, and it includes partitioned biomass components like stem, bark, leaves, and branches (Zhang et al., 2011). Biomass components provide important information for forest management decisions. For example, knowledge of crown biomass aids fuel load assessment and strategies of fire management. Although knowledge of the contribution of each component to AGB is crucial for studying forest growth, it is essential to understand how these components interact with each other (Lambert et al., 2005; Saatchi et al., 2007; He et al., 2013). Meanwhile, the saturation problems resulted from the low penetration of short electromagnetic wave in the forest with high AGB level are the bottleneck problems in forest AGB inversion using remote sensing technology, and accurate estimation of biomass components can improve the estimation saturation points in AGB estimation and reduce the uncertainty of carbon sink estimation, which is the key to quantify carbon stock and plays an important role in modern forest and ecosystem management (Jia et al., 2015; Nie et al., 2017). Therefore, accurate forest total AGB and forest AGB components are crucial. However, accurate estimation of forest total AGB and biomass components is still a challenging task in forestry research at present.

Light Detection and Ranging (LiDAR) is a powerful tool for estimating AGB in forests with fine resolution, which provides detailed three-dimensional information about the forest structure by emitting LiDAR pulses that penetrate the canopy. This is closely related to the spatial heterogeneity of forest carbon content and habitat (Asner et al., 2012; Vaglio Laurin et al., 2016). Total and component AGB were investigated and estimated using LiDAR-derived features in previous research (Næsset and Gobakken, 2008; Tsui et al., 2012; He et al., 2013; Cao et al., 2014). A large number of LiDAR-derived features were demonstrated to be useful of predicting biomass; however, they also revealed that the results were site or species dependent (Zhao et al., 2009; Salas et al., 2010). In parametric regression algorithms, the population assumptions of regression models are usually based on linear relationships. However, it is difficult to fully describe the complex non-linear relationship between forest AGB and LiDAR data using traditional statistical regression methods (Zhao et al., 2019; Torre-Tojal et al., 2022). Several studies found that machine learning algorithms, such as random forest (RF) and support vector regression (SVR) that abandoned the population assumptions that did not represent the heterogeneity of forest stands in parametric regression algorithms, performed better than parametric algorithms in forest AGB estimation (Breidenbach et al., 2010; Vauhkonen et al., 2010; Gleason and Im, 2012; Görgens et al., 2015). Therefore, many studies have applied machine learning algorithms for forest AGB estimation. The performance of machine learning algorithms applied in forest AGB estimation using LiDAR-derived features was not fully explored by far, and their capability for accurate AGB estimation has no agreement (Gleason and Im, 2012). Although the machine learning algorithms show a good performance, the large amount of data constrains the direct use features extracted from LiDAR data as input to the inversion models. For forest AGB estimation using remote sensing observations, a key step is to optimize the optimal features from abundant remote sensing observation (Yang et al., 2017). The importance of each extracted feature was interpreted and ranked by Görgens et al. (Görgens et al., 2015); however, the optimization of feature was not fully explored yet; especially, little was known about component AGB estimation.

For machine learning models to estimate forest AGB, the method of testing model performance is important and involves appropriate validation methods to determine the best predictive model. Leave-one-out cross-validation (LOOCV) and cross-validation are popular for validating the results of forest biophysical parameter estimation. However, how effectively can these two validation methods be used and how different are they and what impact do they have on the estimation results? In our knowledge, it is not addressed yet. According to abovementioned research gap, this study focuses on forest total and component AGB retrieval using height-related features extracted from LiDAR by RF and SVR algorithms. Partial least squares regression (PLSR) algorithm combining the advantages of both principal component analysis (PCA) and multiple linear regression (MLR) was used for ranking and selecting the optimal LiDAR features. LOOCV and cross-validation methods were utilized and compared for results validation. By comparing the two validation methods, we aim to improve the reliability of the model assessment, find the most suitable validation strategy for the inversion model and data in the study area, and improve the efficiency of the utilization of computational resources with the expectation that the model performs well in a variety of data scenarios. The objective is to address the effects of the validation methods and also explore the potential of RF and SVR for forest AGB and component biomass inversion and the potential of PLSR algorithm for selecting the optimal LiDAR-extracted features. To be more concise, we use forest total biomass to describe the forest AGB; biomass components to describe the stem, bark, leave, and branch component biomass; and forest AGBs to describe both total and component biomass. It is expected to provide a valuable reference for the selection of the validation methods in machine learning inverse forest biomass and component biomass studies.

The feature parameters of machine learning models are not limited by dimensionality, and the inversion of forest total biomass using machine learning models has good robustness when there is multicollinearity between feature variables, effectively avoiding the loss of important parameters while maintaining excellent estimation performance (Kotsiantis et al., 2006; Görgens et al., 2015). Therefore, machine learning has been widely used for forest AGB estimation with good accuracy (Cao et al., 2018; Jiang et al., 2022; Li et al., 2022; Torre-Tojal et al., 2022). In the study by Ji et al. (Ji et al., 2022), the accuracy of parametric and non-parametric models for estimating the total forest AGB was compared using SAR data, and the results showed that the non-parametric model gave better estimates than the parametric model, and the non-parametric model was more advantageous than the parametric model in the estimation of the total forest AGB. In our study, two non-parametric models, namely, RF and SVR, were constructed using sample AGB and associated LiDAR remote sensing features for forest total AGB and component AGB inversion. The estimation accuracy of the inverse model was tested by the LOOCV and cross-validation methods.

The results of the comparison between the two machine learning models showed that the R ² value (0.75) of the total biomass estimated by RF was higher than that of SVR when using the LOOCV. This result is consistent with the findings of several studies. For example, the RF and SVR models were compared by Görgens et al. and Kumari and Kumar (Görgens et al., 2015; Kumari and Kumar, 2023), and the result was that the RF model outperformed the SVR model in terms of prediction performance. In the work of Görgens et al. (Görgens et al., 2015), neural network, RF, and SVR models were used to predict stand volume in fast-growing plantation forests, and the RF model produced the best prediction results. The study conducted by Kumari and Kumar (Kumari and Kumar, 2023) compared the predictive potential of SVR and RF algorithms in predicting forest AGB. The result was that the predictive performance of RF is better than that of SVR in this study. In these studies above, the RF algorithm achieved better prediction results compared with the SVR. This may be due to the fact that RF is better at handling non-linear relationships and that the model requires fewer hyperparameters to be adjusted compared with SVR. On the other hand, SVR requires selecting appropriate kernel functions and tuning hyperparameters based on the characteristics of the dataset. However, this does not imply that RF outperforms SVR in all scenarios. Therefore, the selection of machine learning methods for inversion of forest AGB needs to be evaluated and compared on the basis of a combination of various factors. In addition, the results of our study using RF to estimate component biomass showed that leaf biomass was the least correlated with LiDAR data, with an R ² value of 0.54. This result is in agreement with He et al. (He et al., 2013) who used LiDAR data to estimate the summed component AGB of coniferous forests. In their study, a linear regression model was used, and the results also showed weaker performance for leaf AGB estimation but better performance for stem, branch, and total forest biomass estimation. The range of biomass values of leaves is too low, which may be a reason for the weaker relationship.

Height variables extracted from LiDAR data were strongly correlated with the total and component forest AGBs. In this study, height variables change according to the different component of forest AGBs. H ₃₀, H ₄₀, and H _sqt are the optimized height variables for total AGB estimation, and the same variables are selected as optimal features for stem estimation as well. H ₃₀, H _mean, and H ₈ are the selected optimal height variables for bark AGB estimation, whereas, for branches and leaves, they are H _sqt, H _mean, and H ₂₀ and H ₁₀, H ₂₀, and H ₉, respectively. The differences in height variables used for biomass estimation of different components indicate that LiDAR-derived height features have varying explanatory capabilities for biomass composition. It may be related to the extracted height characteristics of LiDAR data to the vertical structure of the forest. In addition, multiple correlations among LiDAR characteristic variables and some overlap among cumulative height percentile variables may also be responsible for the large differences in the relative importance ranking of variables in the component biomass models (Hong et al., 2019). Cao et al. (Cao et al., 2014) estimated total and component biomass in a subtropical forest using small discrete and full waveform airborne LiDAR data. Although the inversion methods that they used were MLR models, their results also confirmed that the height variables extracted from LiDAR data were highly correlated with total forest biomass and component biomass. Forest height features extracted from small-footprint data (Popescu, 2007), small-footprint full-wave form data (Hermosilla et al., 2014a; Hermosilla et al., 2014b), and large-footprint SLICER data (Drake et al., 2002; Lefsky et al., 2005) also explained most of the variability of them for forest structure characteristics.

The comparisons of different validation method are not addressed in other studies, whereas several studies demonstrated that LOOCV performed better especially the limitations of small samples are existent (Zeng et al., 2022; Shi et al., 2023). In this study, the validation difference between LOOCV and cross-validation was compared, and the results revealed that LOOCV showed better accuracy for forest stem, bark, stem, branch, and total AGB estimation but worse accuracy for forest leaf AGB estimation. The accuracy here is related to the value of rRMSE.

Compared with the results of studies using different data sources in the same study area, the height features extracted from LiDAR data outperformed than other data sources. Li et al. (Li et al., 2020) extracted remote sensing features from Landsat8 OLI, Gaofen-1 optical data, and ALOS-1 PALSAR-1SAR to compute forest total biomass at the same test site. They used a fast iterative procedure to optimize input remote sensing features to improve the inversion capability of K-nearest neighbor (KNN) algorithms; the results were validated by LOOCV method; and R² = 0.63 and RMSE = 28.84 Mg/ha were weaker than the R ² and RMSE values for estimated total AGB using RF and SVR and LOOCV validation in this study. Zeng et al. estimated total and component forest AGBs using features extracted from synthetic aperture radar and demonstrated that C-band polarimetric features performed best for forest leaf AGB estimation with R ² = 0.637 and RMSE = 1.27 Mg/ha (Zeng et al., 2022). The estimation of forest total and component AGBs using different data sources revealed the great potential of LiDAR features for accurate estimation.

Although we explored the forest total biomass and biomass component inversion based on optimal LiDAR-derived feature selection with PLSR algorithm and RF and SVR inversion algorithms, there are other machine learning methods that we do not explored in this study. Moreover, the component biomass in this study was calculated by the conversion factors; it may introduce uncertainties for the inversion results; later, the field collected component biomass could be applied in similar study to reduce the uncertainties.

In this study, 56 height-related features extracted from LiDAR were used in RF and SVR algorithms for forest total and component AGB estimation. PLSR algorithm was utilized for ranking and selecting optimal LiDAR-derived features. LOOCV and cross-validation methods were performed to validate the inversion results obtained by RF and SVR. Four cumulative height percentiles (AIH ₃₀, AIH ₄₀, AIH ₂₀, and AIH ₂₅), two height percentiles (H ₈ and H ₆), and four height-related variables (H _mean, H _sqrt, H _mad, and H _curt) are more sensitive LiDAR-derived features for total and component forest height estimation. RF performed better than SVR for both forest total biomass and biomass component estimation. LOOCV showed better accuracy for forest stem, bark, stem, branch, and total AGB estimation but worse accuracy for forest leaf AGB estimation. Note that the difference between the validation using LOOCV and cross-validation is no more than 5%. The features extracted from LiDAR showed a weak performance for leaf AGB estimation when compared with other component AGBs and total AGB estimation. Because only 56 height-related features and only two machine learning methods were applied in this study, future work should focus on exploring more metrices especially derived from full-waveform LiDAR data and more machine learning methods like KNN and Gaussian processes.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

JM: Methodology, Software, Supervision, Validation, Visualization, Writing – original draft. WZ: Conceptualization, Project administration, Resources, Writing – original draft, Writing – review & editing. YJ: Writing – original draft, Writing – review & editing, Validation, Visualization. JH: Validation, Methodology, Formal Analysis, Writing – original draft. GH: Software, Writing – original draft. LW: Visualization, Writing – original draft.