Artificial intelligence-assisted tree diameter-at-breast-height measurement

Abstract Accurate diameter-at-breast-height (DBH) measurement is essential for forest inventory and ecological monitoring. This work evaluates two AI-assisted smartphone frameworks: a text-prompt Segment Anything Model 3 (SAM3) that segments trunks from red-green-blue (RGB) images, and a depth-guided MobileSAM model that uses depth cues to generate geometry-based prompts. The resulting masks are integrated into a geometric pipeline for DBH estimation. The proposed frameworks combine AI-based trunk segmentation with geometry-guided DBH measurement to improve the accuracy, interpretability, and practical efficiency of smartphone-based DBH estimation, with the depth-guided MobileSAM framework enabling near real-time use and the SAM3 for its robustness and accuracy despite of the high computational power requirement. Both methods were tested on 202 planted and 285 natural trees. In planted forests, both achieved high accuracy with no failures (root mean square error, RMSE: 1.2 cm for SAM3, 1.1 cm for MobileSAM). In natural forests, SAM3 remained robust (1.7 cm RMSE), while MobileSAM failed in four cases and showed lower accuracy (2.4 cm RMSE excluding failures). MobileSAM runs in ~0.89 s per image on a smartphone, enabling near real-time use. SAM3 is much slower on CPU (~160 s on 12th Gen Intel(R) Core(TM) i7-1260P) but can be accelerated on GPUs (0.13 s on A100, 3.71 s on T4). Overall, MobileSAM is more practical for on-device deployment, while SAM3 offers greater robustness in complex environments but requires high-performance hardware.