The research workflow behind the open-source harpia project makes that distinction concrete. A team led by Etienne Laliberté described collecting wide-area imagery and then sending a drone back to individual canopy targets for much closer photographs. Their 2026 Methods in Ecology and Evolution paper reports field results from tropical sites; the repository documents how the targeting and flight-plan generation work. The useful engineering lesson extends beyond tropical botany: a coordinate is only the beginning of a safe and reproducible imaging mission.
A map and a botanical photograph answer different questions
An initial mapping flight produces the spatial context. A nadir survey can support an orthomosaic, a digital surface model (DSM), and the outlines or centers of target crowns. That is enough to locate candidates and estimate their elevations. It is not necessarily enough to recognize leaves, flowers or bark on a tall tree.
The second pass changes the imaging scale. In the published experiments, the researchers used a DJI Mavic 3 Enterprise to acquire close-ups approximately 6 m above target foliage, reporting imagery at roughly 0.4 mm per pixel under their field conditions. They report that this detail helped botanical experts identify many of the sampled trees. Those are observations from particular sites and workflows, not a guarantee that every species or crown position can be identified remotely. See the paper and Mila's account of the project.
The distinction matters when choosing a survey design. A wider camera view reduces flight time and gives better geographic context; close-ups demand more positions, more careful height control, and a stable view of the target. Optimizing one capture mode does not eliminate the need for the other.
Geographic coordinates have to survive several transformations
Harpia accepts target points or polygons in GeoPackage or Shapefile form, together with an elevation-bearing DSM GeoTIFF. Its documentation specifies that the feature layer, optional area of interest and DSM share a projected coordinate reference system. It then generates geographic waypoint coordinates and builds mission files for DJI Pilot 2 using KML and Wayline Markup Language (WPML), packaged in KMZ.
Each hand-off has a failure mode. A projected eastings/northings pair interpreted as latitude/longitude is obvious corruption, but a subtler mismatch between datums, elevation references, image footprints and polygon centers may still yield plausible-looking points. The repository specifically requires DSM elevation values in meters referenced to an ellipsoidal height model. Ground elevation, canopy-surface elevation, and the aircraft's commanded height must not be treated as interchangeable.
A reproducible flight package therefore needs the source imagery, DSM version, coordinate reference systems, feature-selection method, vertical reference and mission-export settings recorded together. Reviewing the waypoint plot without reviewing the elevation model leaves a major part of the plan unexamined.
The shortest tour may not be the flyable tour
Once crowns have been selected, Harpia estimates target heights from the DSM, orders visits using a traveling-salesperson-style route calculation, and inserts checkpoints intended to take the aircraft above obstructions on its way between targets. The repository's waypoint-generation specification describes both target waypoints and clearance checkpoints in the intermediate CSV.
The distinction is more important in a forest than on an open field. Two nearby crowns in plan view may be separated by a taller emergent tree, a steep canopy edge or a radio-shadowed corridor. A short straight line between their centroids can be a poor operational route. The minimum path over a flat map is not necessarily the minimum safe path through three-dimensional vegetation.
Nor is a DSM a live obstacle-avoidance system. Its resolution, acquisition date and occluded surfaces limit what it represents. Branches move in wind, canopy gaps change, and thin obstacles may not appear in the surface model. The generated path remains a plan requiring site assessment and appropriate aircraft safeguards, rather than a proof that every segment is clear.
RTK position quality depends on the communications path
Close-up photography requires the aircraft to return to small targets. The researchers used real-time kinematic (RTK) positioning to reduce geographic uncertainty, and the paper reports target geolocation errors below 13 cm in its tested workflow. That figure describes the authors' measurement conditions; it does not certify the accuracy of every subsequent mission or the position of every pixel.
The Harpia README documents a configurable "touch-sky" procedure: the aircraft periodically climbs higher to improve its connection and obtain updated RTK corrections before returning toward low-level targets. It is optional rather than the default. The software allows interval and climb-height parameters, but an available setting is not permission to climb above a site's altitude limit.
This is a useful operational trade-off. Reconnecting high above the canopy may protect the correction link, but climbs consume time and battery and introduce their own airspace and separation concerns. The researchers' reported link performance at some sites, including distances of kilometers, should not be translated into a universal range claim. Terrain, vegetation, antennas and radio conditions can change the link budget sharply.
The research results are a baseline, not a flight guarantee
The paper reports that the team could image on the order of 300 selected crowns per day at its field sites. It also reports high-resolution imagery and positional performance under the tested conditions. These are valuable feasibility results because they combine capture geometry, field throughput and a scientific use case. They are not controlled benchmarks against every competing workflow.
The public Harpia software presently documents both Mavic 3 Enterprise and Matrice 4 Enterprise mission output. That software compatibility should not be confused with a claim that the paper's field measurements were independently replicated on both aircraft. Likewise, an automated photo route does not automatically validate the crown annotations, botanical classifications or downstream machine-learning results.
A rigorous follow-up study would disclose the denominator behind a success rate: how many proposed crowns were rejected during preselection; how many missions required intervention; how many photographs were obscured or out of focus; and whether the same tree could be found consistently across days. Accuracy should be checked against surveyed controls and representative ground observations rather than against RTK status alone.
Make the mission auditable before making it autonomous
The most transferable part of this workflow is its separation of decisions. First identify a crown on an areal survey. Then derive a target and height from a specific DSM. Then plan travel and obstacle clearance. Finally record whether each image actually captured the intended target.
The repository includes optional filters for area of interest, minimum crown area, previously visited points, reference polygons and local DSM relief. Those filters make selection reproducible if their inputs and thresholds are retained. Its CSV schema also distinguishes target waypoints from checkpoints and carries identifiers, height fields and ordering information. These are practical building blocks for comparing an intended visit with what happened in flight.
An operator adapting the approach should keep a per-target record linking the feature identifier, map version, planned height, RTK quality or correction state, photo timestamp, camera pose, and any manual override. The project does not claim to supply all of that validation automatically. It is an audit design implied by the workflow, not a feature that should be credited to the software without testing.
Where the workflow will need to change
The model assumes that the planned surface is a useful approximation of the objects being photographed. It will be less reliable where vegetation moves substantially, overlapping crowns hide the target, new growth has altered the canopy, or the DSM is too coarse to represent the relevant clearance. Species recognition can still fail even when the flight path succeeds, because botanical features may be hidden, seasonal or absent.
The first field question is therefore not "Can the optimizer visit every point?" It is "Can we trust the coordinates, the vertical model and the line of sight well enough to approach each point?" The answer must include local flight rules, radio conditions, aircraft performance and a plan for targets that cannot be approached safely.
Harpia's open implementation and the published field study make the architecture inspectable. They do not remove the need to validate the particular forest and aircraft. The strongest use of automation here is to produce a testable mission plan whose assumptions remain visible, rather than to turn a list of GPS points into an unquestioned flight path.