In fall, New Mexico’s Middle Rio Grande wetlands become packed with wintering waterfowl.
In the past, surveys involved pilots flying low enough so that birds could be counted from the plane’s windows, but now, drones are doing the job.
Deep learning applied to the aerial images returned a number of 20,000-plus ducks.
This was not accurate, and when the algorithm executed a correction pass, phantom birds were eliminated and real ducks overlooked in vegetation were added.
Why counting birds from planes fails
Biologists monitoring the migration flyways of North America have relied on aircraft surveying for decades.
The method is not ideal, as the planes have to fly at risky low altitudes over open water while humans try to estimate bird numbers at high speed and in glaring light.
Accurate counts are impossible across large, dense flocks.
The estimates gathered during a single pass often vary notably between different observers surveying the same marsh area, so new strategies were needed for more accurate, safer data logging.
A fixed-wing drone was deployed over the Rio Grande
Uncrewed aerial observations mean physical risks are eliminated.
Researchers were able to collect photographic records of the marshland habitat by deploying a fixed-wing drone over the Middle Rio Grande wetlands of New Mexico.
The craft was directed to key wintering sites.
While flying systematic transects, high-resolution images were captured in overlapping grids, showing thousands of birds resting on the mudflats or sitting on the water.
More than 20,000 ducks were logged in the wetland ecosystem.
Computer vision and phantom bird errors
It takes weeks of intense labor to manually count thousands of individual birds in photographs.
The advancement of deep-learning models means more rapid detection by applying visual pattern recognition; in this case, the parameters were loaded to target waterfowl.
Initial computer runs came back with errors.
Complex natural background textures caused systemic faults as the algorithm identified reflected sunlight, rocks, clumps of grass, and debris as ducks.
These were all false positives that invented hundreds of birds on the water that did not exist.
Teaching deep learning to spot missed ducks
Automated detection also suffers from false negatives when real birds blend into their surroundings.
Waterfowl packing tightly together overlap each other, obscuring clear individual outlines, while muted winter plumage hides ducks sitting against dark mud, cattails, or deep shadows.
Researchers implemented a multi-stage review process to audit and refine the neural network’s output.
Instead of modifying image detections frame by frame, the researchers used platform-specific confusion matrices to statistically adjust for missed birds and false detections. These calibrated counts were then combined with design-based ratio estimators to yield the final abundance figure.
What the corrected survey data proves
Combining aerial drone surveys with audited computer vision yields verifiable, repeatable population estimates.
Correcting both the false positives and false negatives stabilizes total population estimates across variable marsh conditions, and high-resolution photo archives allow independent researchers to re-examine raw data years later.
Standardized drone flights reduce noise disturbance to resting migratory flocks.
Wildlife agencies can now send out automated monitoring across vast flyways without any risk to human life or deployment of loud motorized boats or planes flying at low altitudes.
Replacing human flight counts with uncrewed aerial drone systems marks a fundamental shift in modern wildlife management.
Although deep-learning algorithms hallucinate phantom targets and miss obscured animals at first, secondary correction passes successfully eliminate both sources of errors.
Drones eliminate dangerous pilot hazards while delivering precise imagery across remote wetland ecosystems.
This practical combination of fixed-wing robotics and refined computer vision provides conservation agencies with a scalable, safer blueprint for tracking migratory bird populations nationwide.
All the results of the study can be read here: Stewart, D. R., Converse, R., Sesnie, S. E., Butler, M. J., Moon, J. A., Sanspree, C. R., … & Lippitt, C. (2026). A design‐based framework for estimating wildlife density from uncrewed aerial systems imagery. The Journal of Wildlife Management, e70271.
Read the whole thing?
Get the week's signal, not the noise
Our sharpest reporting on energy, climate and nature — free, once a week.
