Interpretable and probabilistic forest stock volume estimation using UAV-LiDAR and NGBoost

Forest stock volume (FSV) is a key indicator for forest resource assessment, carbon stock accounting, and forest management decision-making. UAV-LiDAR can provide detailed three-dimensional information to describe forest structure parameters. However, most existing machine learning methods for FSV estimation focus primarily on point prediction and have difficulty simultaneously providing prediction uncertainty and interpretable model diagnostics. To address this issue, this study developed an interpretable probabilistic modeling framework using forest plots in Nanning, Guangxi Zhuang Autonomous Region, China. Based on UAV-LiDAR data and 30 field plots of 30 m × 20 m collected in 2022, eight 10 m × 10 m subplot-level samples, including partially overlapping samples, were generated from each original plot, yielding 240 samples. To reduce spatial data leakage, the dataset was split at the original-plot level. The 192 subplots derived from 24 original plots were used as the training set, within which cross-validation was conducted for feature selection and hyperparameter optimization, whereas the 48 subplots derived from the remaining six plots were reserved as an independent test set. The proposed framework integrates Chatterjee-Boruta feature selection, Natural Gradient Boosting (NGBoost) probabilistic prediction, and dual SHapley Additive exPlanations (SHAP) interpretation. First, key structural features were selected from 56 LiDAR-derived variables. NGBoost was then used to predict the Log-Normal conditional distribution of FSV. Finally, the structural drivers of the location parameter μ and the scale parameter σ were interpreted separately. Chatterjee-Boruta retained 22 LiDAR-based features, mainly including accumulated height percentile metrics, height percentile metrics, and height statistical metrics, with representative variables such as elev_AIH_25th, elev_p_5th, and elev_crr. Results on the independent test set showed that CB-NGBoost achieved a balanced performance in both point prediction accuracy and probabilistic prediction quality, with an R² of 0.74, RMSE of 71.24 m³·ha⁻¹, MAE of 52.81 m³·ha⁻¹, and rRMSE of 39.70%. In terms of probabilistic prediction, CB-NGBoost obtained the lowest Continuous Ranked Probability Score (CRPS, 39.79 m³·ha⁻¹), outperforming the other NGBoost combinations and QRF combinations. This indicates that its predictive distribution was more consistent with the observed values overall. The dual SHAP results showed that AIH metrics were the primary source of structural information affecting the central tendency parameter μ of FSV prediction, whereas height statistical metrics had a more pronounced influence on the scale parameter σ. This suggests that the LiDAR structural factors driving FSV prediction values and those driving aleatoric uncertainty are not entirely identical. This study provides an interpretable probabilistic extension to conventional UAV-LiDAR-based FSV point prediction, providing uncertainty-aware FSV estimates that may support forest resource assessment and carbon-stock accounting. However, the current model mainly characterizes aleatoric uncertainty and does not explicitly incorporate epistemic uncertainty, resulting in under-calibrated and overly narrow prediction intervals. Future studies should further address this limitation by incorporating model ensembles or Bayesian methods to represent model-related uncertainty, using conformal prediction to improve interval calibration, and validating the generalizability of the proposed framework across multiple regions and forest stand types.