Can a Drone Count Every Tree in a Community Forest?

A crown is not the same as a stem

A drone can capture detailed images of a forest canopy. It cannot see every tree beneath it.

At Shree Manakamana Community Forest in Kawasoti, GeoSutra used low-altitude imagery, field plots and a locally retrained crown-detection model to map visible tree crowns. The project page reports 6,543 detections and a model score of 0.73 mAP at an intersection-over-union threshold of 0.50.[1]

Those figures describe parts of a mapping exercise. They do not establish a complete count of trees in the forest.

In a high-resolution image, a model can identify shapes that look like individual tree crowns. The resulting map can help show where visible crowns are concentrated and support further questions about forest structure. But several trees may overlap in the canopy and appear as one crown. Smaller or suppressed trees beneath the canopy may not be visible at all. GeoSutra’s own project page warns that detections are not a tree census.[1]

What a crown map does not measure

A crown map also does not directly provide measurements such as species, stem diameter or tree height. Forest inventories often collect those attributes through field observation and measurement; the right set depends on the inventory’s purpose.[3] A detection layer can be useful alongside that work, but it cannot stand in for variables the imagery and model do not measure.

What does a score of 0.73 mean?

GeoSutra reports a score of 0.73 mAP at an intersection-over-union threshold of 0.50, comparing model detections with hand-annotated reference crowns in held-out plots.[1] Mean average precision summarizes the relationship between precision and recall across detection confidence thresholds. It is not a statement that the model counted 73% of all trees correctly.[4]

The published project record says this evaluation applies to the Manakamana site and season. It does not establish how the model will perform in a different forest, season or sensor setup.[1][2]

One count is still unresolved

The project page labels the total of 6,543 detections provisional. Its four published block counts add up to 6,309, leaving a difference of 234 that GeoSutra has not explained.[1][2]

That gap should remain visible until it is reconciled against the source records. Reporting both figures and the unresolved difference is more useful than presenting a single total as settled.

Use the map as one part of the inventory

Aerial imagery can provide a detailed view of visible crowns across a mapped area. Field measurements and local records can answer other questions, including those involving stems or attributes the camera cannot observe. If a result will guide management, the team should state how observations were selected, what was measured, and which claims the evidence supports.

For Manakamana, the public project record describes a crown map and stand statistics prepared for community forest planning.[1] That is a project report, not an independent audit of the underlying annotations or validation data. Readers should treat the reported results within that scope.

A useful inventory does not begin and end with a model’s count. It combines imagery, field evidence and local needs, and makes clear where each source stops being reliable.