Cheap Drone Photogrammetry Nails Canopy Height but Struggles to Count Individual Trees

Drone photogrammetry has become one of the most attractive tools in forest monitoring because it promises tree-level information at a fraction of the cost of laser scanning. A new study from Alabama, however, delivers a sobering and unusually honest accounting of what the technique can and cannot do. Swati Singh and Lana L. Narine systematically varied the processing settings of Structure-from-Motion, or SfM, reconstruction across thirteen forest plots and found that while the resulting canopy height models were remarkably accurate, the automatic detection and counting of individual trees fell far short of what would be needed for operational inventory.

The research, published in Discover Sensors, focused on loblolly pine (Pinus taeda) stands in two Alabama counties with deliberately contrasting character. Eight plots in Washington County in the southwest are managed, even-aged pine plantations on relatively flat terrain. Five plots in Cullman County in north central Alabama are mixed pine-hardwood stands dominated by oak and sweetgum, with greater crown overlap and more pronounced topographic variation, with elevations ranging from roughly 153 to 286 meters compared with 20 to 131 meters in Washington County. Each plot covered 1.21 hectares, or three acres, and field crews inventoried trees following Forest Health Monitoring and USDA Forest Inventory and Analysis protocols, tagging stems and recording GPS positions that later served as ground truth.

Data were collected in June 2023 with a DJI Mavic 3 M carrying a 20-megapixel RGB camera and an integrated multispectral sensor. Flights ran under clear skies between 10:00 and 12:00 local time to minimize shadow displacement, at 180 feet above ground with 80 percent forward and 75 percent side overlap, an airspeed of 3 meters per second, and a ground sampling distance of about 2.5 centimeters. By holding every acquisition variable constant, the authors could isolate the effect of processing choices, which was the whole point: SfM outputs are notoriously sensitive to how the reconstruction is configured, and a single default setting rarely transfers across sites.

The core experiment was a full 5 by 4 factorial design in Agisoft Metashape Professional. Dense-cloud quality was varied across five levels, from Lowest to Ultra High, and depth-map filtering across four levels, from Disabled to Aggressive, producing twenty parameter combinations. For each combination the researchers measured point density, mean normalized height, height standard deviation, and surface roughness, then tested the effects of quality, filter, and their interaction with two-way ANOVA. Reconstruction quality significantly affected every metric, while filtering significantly affected point density and surface roughness but not mean height or height variability. A significant quality-by-filter interaction appeared only for point density, showing that filtering intensity modulates how density scales with reconstruction resolution.

The post-hoc comparisons told a consistent story. Within every quality level, the Aggressive filter produced lower point densities than Disabled, Mild, or Moderate, and the Disabled and Mild filters yielded higher surface roughness, consistent with stronger filtering suppressing small-scale canopy texture. Mean height showed only weak dependence on quality, with differences among the upper quality levels below 0.4 meters. Balancing density gains against roughness control through a Pareto screen and a composite score, the team selected a single operational configuration: Ultra High reconstruction quality with depth-map filtering disabled. This setting preserves fine-scale canopy texture and treetop detail while avoiding the peak erosion that aggressive filtering can cause.

With that configuration locked in, the canopy height models delivered impressive vertical accuracy. In Washington County, the CHMs matched field-measured heights with a coefficient of determination of 0.96, a root mean square error of 0.44 meters, and a bias of only minus 0.01 meters. In Cullman County, agreement was weaker, with R-squared of 0.44, RMSE of 1.14 meters, and bias of minus 0.09 meters, reflecting underestimation of canopy peaks in the mixed, sloped terrain. Pooled across all plots, performance reached R-squared of 0.98, RMSE of 0.54 meters, and bias of minus 0.01 meters. The authors attribute the site contrast to stand structure and terrain: homogeneous, even-aged pine on flat ground supports consistent reconstruction, whereas overlapping crowns and topographic relief introduce irregular canopy surfaces and reduce point-cloud continuity.

Then came the individual tree detection results, and they were stark. Using a fixed workflow of adaptive local-maxima treetop detection and Dalponte2016 crown segmentation in the lidR package, with detections matched to field stems at a 3-meter horizontal tolerance and a plus-or-minus 10 percent height tolerance, the workflow achieved a precision of 0.025, recall of 0.286, and F1 of 0.047 in the even-aged Washington plantations, and a precision of 0.032, recall of 0.118, and F1 of 0.051 in the mixed Cullman stands. Pooled F1 was 0.047. In plain terms, the algorithm missed most real trees and flagged many false ones. Sensitivity analyses at 2-meter and 4-meter matching radii showed the expected trade-off: tighter matching lowered F1, looser matching raised it but inflated commission error.

The authors diagnose the failure modes precisely. Over-segmentation arose because irregular canopy surfaces and overlapping crowns generated multiple local maxima within single tree crowns, while suppressed and subdominant trees beneath dominant crowns were missed due to canopy occlusion. Washington’s even canopy produced smoother CHM surfaces and fewer spurious peaks, hence its higher recall, whereas the combination of mixed-species canopy and local terrain variation in Cullman drove both over-segmentation and omission. The fundamental constraint is physical rather than algorithmic: SfM reconstructs the illuminated canopy surface, so anything hidden beneath it simply does not exist in the data, and terrain reconstruction under dense canopy carries uncertainty that can propagate into height estimates.

The practical guidance that emerges is nonetheless valuable. For applications needing accurate crown mapping, such as inventory, density estimation, or gap analysis, airborne or UAV lidar remains the preferred technology because laser pulses sample through small canopy gaps and resolve sub-canopy structure. RGB-SfM, by contrast, is well suited to canopy height modeling, relative change detection, and low-cost baseline monitoring where lidar is unavailable. The study also emphasizes that differences in reconstruction settings can introduce variability comparable to ecological differences among sites, which argues for standardized reporting of SfM parameters, and for pairing height metrics with full detection counts of true positives, false positives, and false negatives to enable honest cross-study comparison.

Future directions proposed by the authors include fusing SfM with lidar to combine surface detail with canopy penetration, applying machine-learning crown segmentation that learns from tree shape, and running repeat surveys to track canopy change over time. The processing scripts are publicly available through the Open Science Framework, making the workflow reproducible and transferable. In an era when cheap drones tempt foresters to skip expensive laser scanning, this study is a useful corrective: with careful parameterization, consumer-grade photogrammetry can map forest height with sub-meter accuracy, but counting individual trees in complex canopies still demands a sensor that can see between the leaves.

Subject of Research: Optimizing Structure-from-Motion processing parameters for RGB-only individual tree detection in loblolly pine and mixed pine-hardwood forests

Article Title: Optimizing SfM parameters for RGB-only individual-tree detection in loblolly pine (Pinus taeda L.) and mixed pine-hardwood stands

Article References: Singh, S., & Narine, L. L. (2026). Optimizing SfM parameters for RGB-only individual-tree detection in loblolly pine (Pinus taeda L.) and mixed pine-hardwood stands. Discover Sensors, 2(1), Article 42. https://doi.org/10.1007/s44397-026-00076-8

Image Credits: AI Generated

DOI: 10.1007/s44397-026-00076-8

Keywords: UAV photogrammetry, Structure-from-Motion, individual tree detection, canopy height model, loblolly pine, lidar, remote sensing, forest inventory, Agisoft Metashape, depth-map filtering, point cloud density, Alabama forests