Purpose Accurate estimation of Leaf Area Density (LAD) at fine spatial scales remains challenging across a wide range of canopy architectures and foliage configurations, due to occlusion, variable sampling density, and complex laser beam–canopy interactions. This study introduces a beam-aware learning framework to enrich Mobile Terrestrial Laser Scanning (MLTS) derived point clouds point clouds with sensor-centric descriptors, enabling fine-scale canopy density characterisation. Methods A learning framework was developed using physically based MLTS simulations of orchard canopies. Point clouds were augmented with beam-scale descriptors derived from curved voxel analysis and sensor geometry. Classical machine learning regressors and a deep learning model were trained on local canopy regions extracted at multiple spatial scales and evaluated using cross-tree and stratified cross-validation schemes. Results The inclusion of beam-informed descriptors consistently improved LAD estimation accuracy compared with Cartesian-only representations. Under cross-tree validation, the proposed deep learning model achieved <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textbf{MAE}=0.1\)</EquationSource> </InlineEquation>, outperforming classical baselines such as Support Vector Regression, Random Forest, and Gradient Boosting, whose best performance reached <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\textbf{MAE}\,\varvec{\approx 0.2}\)</EquationSource> </InlineEquation>. Under stratified conditions, the model attained <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\varvec{R}^2\varvec{\approx 0.95}\)</EquationSource> </InlineEquation>. Feature analysis identified local gap fraction and beam interaction metrics as the most informative attributes. Conclusion Extending point-level representations beyond Cartesian coordinates by incorporating beam- and sensor-centric descriptors significantly improves voxel-level LAD estimation at local scale, and supports Leaf Area Index (LAI) estimation at tree scale in orchards. The proposed framework provides a scalable solution for fine-scale canopy characterisation, supporting MLTS-based monitoring in precision agriculture and agricultural robotics.

A beam-informed framework for Leaf Area Density estimation from Mobile Terrestrial Laser Scanning

期刊论文
发表日期：

2026年7月29日

查看全文

Harold Murcia,
Simon Lacroix

摘要

Purpose

Accurate estimation of Leaf Area Density (LAD) at fine spatial scales remains challenging across a wide range of canopy architectures and foliage configurations, due to occlusion, variable sampling density, and complex laser beam–canopy interactions. This study introduces a beam-aware learning framework to enrich Mobile Terrestrial Laser Scanning (MLTS) derived point clouds point clouds with sensor-centric descriptors, enabling fine-scale canopy density characterisation.

Methods

A learning framework was developed using physically based MLTS simulations of orchard canopies. Point clouds were augmented with beam-scale descriptors derived from curved voxel analysis and sensor geometry. Classical machine learning regressors and a deep learning model were trained on local canopy regions extracted at multiple spatial scales and evaluated using cross-tree and stratified cross-validation schemes.

Results

The inclusion of beam-informed descriptors consistently improved LAD estimation accuracy compared with Cartesian-only representations. Under cross-tree validation, the proposed deep learning model achieved \(\textbf{MAE}=0.1\) , outperforming classical baselines such as Support Vector Regression, Random Forest, and Gradient Boosting, whose best performance reached \(\textbf{MAE}\,\varvec{\approx 0.2}\) . Under stratified conditions, the model attained \(\varvec{R}^2\varvec{\approx 0.95}\) . Feature analysis identified local gap fraction and beam interaction metrics as the most informative attributes.

Conclusion

Extending point-level representations beyond Cartesian coordinates by incorporating beam- and sensor-centric descriptors significantly improves voxel-level LAD estimation at local scale, and supports Leaf Area Index (LAI) estimation at tree scale in orchards. The proposed framework provides a scalable solution for fine-scale canopy characterisation, supporting MLTS-based monitoring in precision agriculture and agricultural robotics.