A multisource object-based upscaling framework for growing stock volume estimation integrating TLS, ALS, and orthophoto

Abstract

Accurate estimation of individual tree volume at large spatial scales remains a major challenge in forest biomass assessment. Existing approaches struggle to reconcile the high structural accuracy of Terrestrial Laser Scanning (TLS) with the extensive spatial coverage of Airborne Laser Scanning (ALS) and satellite imagery. Conventional upscaling methods, which are primarily based on pixel- or grid-level representations, often fail to preserve tree-level structural integrity, leading to reduced accuracy in heterogeneous forests.To address this limitation, this study proposes a multisource object-based upscaling framework that integrates TLS-derived reference volumes used as training data with ALS structural metrics and high-resolution orthophoto features. The framework is guided by the hypothesis that modelling units aligned with structurally representative individual tree crowns can better represent forest structure than conventional spatial units. Accordingly, the central research question is whether object-level, crown-based representations improve large-scale tree volume estimation compared to pixel- and grid-based approaches. We hypothesise that preserving these crown structures enables more accurate and ecologically consistent upscaling.The framework was developed using TLS-derived volumes from 674 trees, aggregated into 320 crown objects delineated from ALS-derived canopy height models and orthophotos. To evaluate its effectiveness, multiple upscaling configurations were evaluated, including Sentinel-2 only, PlanetScope only, hybrid optical–LiDAR models, fixed-resolution grid-based LiDAR–orthophoto models, and two proposed object-based frameworks that combine LiDAR and orthophoto data with and without species stratification. These object-based frameworks represent the central methodological contribution of this study. A three-stage feature selection pipeline integrating Random Forest importance, correlation pruning, and recursive elimination was followed by five machine-learning regression algorithms (Linear, Support Vector Regression, Random Forest, Gradient Boosted Regression, Extreme Gradient Boosting).The object-based LiDAR-orthophoto model, using Support Vector Regression, achieved the highest predictive performance (R² = 0.87, RMSE = 0.07 m³, rRMSE = 23.8%) and outperformed all pixel- and grid-based alternatives for a heterogeneous set of tree species. Stratified modelling further improved accuracy for structurally homogeneous species such as Eucalyptus (R² = 0.93, RMSE = 0.04 m³). These findings demonstrate that preserving crown-level structure is critical for accurate upscaling, and that integrating three-dimensional structural information with object-based representations provides a robust and scalable framework for tree volume estimation that preserves individual-tree structural variability, which is essential for applications such as forest inventory, carbon accounting, and ecosystem assessment.