What question did this study set out to answer?

The research aims to develop an accurate and interpretable framework for estimating aboveground biomass (AGB) using UAV data and machine learning.

March 14, 2026Open Access

Fine-scale aboveground biomass estimation of restored vegetation in tailings ponds: integrating UAV multi-source data and explainable machine learning

Key Points

The research aims to develop an accurate and interpretable framework for estimating aboveground biomass (AGB) using UAV data and machine learning.
Constructed a framework using UAV LiDAR and hyperspectral data.
Selected 38 key variables using the Boruta algorithm from 128 features.
Built multiple machine learning models including RF, XGBoost, SVM, and KNN, with RF yielding the highest accuracy.
Performed SHAP analysis to assess feature importance across AGB gradients.
The RF model achieved an accuracy of R2 = 0.86 and RMSE = 29.22 Mg/ha.
Independent validation confirmed robustness with R2 = 0.78.
AGB estimates ranged from 83.31 Mg/ha to 396.43 Mg/ha, with the highest values in northern regions.
Populus made up 79.8% of total AGB, while Robinia had a higher carbon stock capacity per area.

Abstract

• Constructs an interpretable ML framework fusing UAV LiDAR and hyperspectral data. • Boruta-RF model achieves high accuracy (R 2 = 0.86) by overcoming spectral saturation. • LiDAR height metrics (H10, H20, H25) critically mitigate signal saturation effects. • SHAP analysis reveals feature shift from texture to height metrics in high-AGB zones. Accurate and automated quantification of aboveground biomass (AGB) is a prerequisite for the precision management and digital evaluation of ecological restoration in complex terrains, such as tailings ponds. However, traditional optical remote sensing often suffers from signal saturation in dense vegetation and lacks interpretability in modeling. To handle these challenges, this study created an AGB inversion framework integrating UAV multi-source data and explainable machine learning (ML). Using the Boruta algorithm, 38 key variables were selected from an initial set of 128 features extracted from hyperspectral imagery and LiDAR data. These features were utilized to construct RF, XGBoost, SVM, and KNN models, with the RF model achieving the highest accuracy with R 2 = 0.86 and RMSE = 29.22 Mg/ha, and superior stability. Furthermore, independent validation using 24 external samples confirmed the model’s robustness (R 2 = 0.78). SHapley Additive exPlanations (SHAP) analysis revealed that while spectral vegetation indices were the dominant contributors overall, feature importance shifted across AGB gradients. Texture features were critical drivers in low AGB regions, whereas LiDAR canopy height metrics played a decisive role in high AGB areas, effectively mitigating spectral saturation. The estimated AGB exhibited marked spatial heterogeneity, ranging from 83.31 Mg/ha to 396.43 Mg/ha, with the highest AGB clustered in the northern regions. Furthermore, while Populus constituted 79.8% of the total AGB, Robinia demonstrated a superior aboveground carbon stock capacity per unit area. This framework offers a robust, interpretable approach for fine-scale AGB estimating, supporting green mine development and carbon neutrality goals.

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper