A consumer-facing application for radiance field-based 3D reconstruction enables accessible measurement of urban tree structure

Sruthi M. Krishna Moorthy, Maren Lechner, Henry Cerbone, Bhavya Palgudi, Graham Taylor, Rob Salguero-Gomez

Urban forests are key in mitigating the effects of climate change and urbanisation on human well-being. Yet, evaluating and managing these benefits requires accurate, scalable monitoring of trees. Laser scanners operated from ground-based and airborne platforms have enabled urban trees to be monitored in unprecedented structural detail and across large spatial extents, respectively. However, important trade-offs remain among cost, spatial coverage and the level of structural detail captured. Recent advances in radiance-field-based continuous 3D reconstruction offer a more accessible alternative, while their integration into consumer-facing smartphone applications has opened new opportunities for participatory monitoring via citizen science. However, these approaches remain insufficiently validated for urban trees. We evaluated radiance-field-based 3D tree reconstructions generated from smartphone videos using the consumer-facing Luma AI platform, comparing them with terrestrial laser scanning (TLS) reference data across leaf-on and leaf-off seasons and a gradient of surrounding vegetation complexity. Seventy-eight trees across fourteen urban plots of size 20 m by 20 m were surveyed under leaf-on and leaf-off conditions. Luma AI produced usable reconstructions for 63 trees in summer and 66 in winter, with no significant seasonal difference in failure rate (19.2% versus 15.4%; McNemar p = 0.629). Tree height showed strong agreement with TLS in both summer and winter (CCC = 0.887 and 0.866, respectively), while crown-area agreement was lower, particularly during winter (CCC = 0.795 and 0.643). Fine branches, which were occluded in summer but exposed during winter data collection, were more difficult to reconstruct, resulting in winter trees having higher precision but lower recall and overall lower reconstruction fidelity than summer trees (F1 = 0.366 ± 0.220 versus 0.420 ± 0.232). Nevertheless, this loss of geometric completeness did not prevent the extraction of useful whole-tree structural metrics. Reconstruction fidelity declined with increasing tree density and total crown area during summer (Spearman’s ρ = −0.677 and −0.692), when foliage increased occlusion. Consumer-facing applications enabling continuous-scene reconstruction methods therefore offer considerable potential for large-scale participatory monitoring of trees in open urban landscapes and could help democratise structural monitoring in other sparsely wooded ecosystems.