Snake researchers review fewer images with MegaDetector

MegaDetector with Timelapse2 [1]

MegaDetector, based on YOLOv5. [1] [2]

MegaDetector 5A [1]

MegaDetector contributors [1]

MegaDetector detected animals. Timelapse2 supported human review and species identification. [1] [2]

Unadapted MegaDetector 5A ran on an Nvidia A100 GPU on UF HiPerGator. The reported comparison used a confidence threshold of 0.05. [1] [2]

Time-lapse photographs were the input. Detection scores and locations selected images for human species identification. [1] [2]

The model filtered images for review. It did not identify snake species or initiate conservation action. [1] [2]

Researchers collected images, compared manual review with MegaDetector filtering followed by human review, and reconciled detections from both methods. [1]

A researcher or herpetologist identified species. The same reviewer supported both workflows, and the paper does not report randomization or blinding. [1]

AI was part of the measured image-filtering workflow. Human review supplied species labels and the comparison data. [1]

MegaDetector filtered images by detecting animals. People identified species. The same image set and reviewer supported a comparison, but hybrid human review time was estimated and overall detection was lower. [1]

AIBD-0583-outcome-research [1]

Miller and colleagues evaluated MegaDetector 5A on 1,485,924 time-lapse images from 42 cameras at three South Florida sites. At a detection threshold of 0.05, 88,863 images remained for human review, a reported 94.0% reduction. Manual review took 265 hours. Computer processing took 27.7 hours, and the authors estimated a 43.5-hour hybrid workflow by adding human review time calculated from the manual review rate. About 16 hours of that total represent estimated human review, not separately timed hybrid labor. MegaDetector located animals; people assigned snake species in Timelapse2. The hybrid method found 217 independent snake events, compared with 378 manually and 447 across both methods. Snake-image recall was 17.1% for the hybrid method and 96.2% manually. Results differed by species: the hybrid method found five of six Burmese-python events and more North American racer events than manual review. Neither method found every event. The study documents reduced review work with detection losses. It does not document a management decision, population improvement, or net financial saving. [1]

Improved Organizational Efficiency [1]

Reported research result with explicit limitations. [1]

Participating research organizations. Wider conservation benefits remain unmeasured. [1]

The research team that processed the South Florida snake images. [1]

The study does not count people who benefited. Cameras, images, species, and detection events are study units. [1]

Performance differed across snake species and sizes. The hybrid workflow retained five of six Burmese-python events but missed the ring-necked-snake event that manual review found. [1]

The study does not establish how long the benefit persisted after the evaluation. [1]

The result concerns the intended use of AI to support wildlife-image analysis. [1]

Reduced image-review work came with fewer total snake detections. The hybrid workflow found some events that manual review missed, and neither method established complete ground truth. [1]

AIBD-0583-evaluation-miller-2025 [1]

Comparison of manual and AI-assisted review of the same image dataset, using the same reviewer. The paper does not report a randomized or blinded workflow comparison. [1]

The authors timed manual review and computer processing, estimated hybrid human review from the baseline review rate, and compared detections after combining findings from both methods. [1]

42 cameras at three South Florida sites produced 1,485,924 images between June 20 and July 18, 2022. [1]

Of 1,485,924 images, 88,863 remained for human review. Manual review took 265 hours; the hybrid workflow was estimated at 43.5 hours. Hybrid and manual review found 217 and 378 independent snake events, respectively. [1]

Manual review of the same time-lapse image dataset. [1]

At threshold 0.05, the hybrid workflow left 88,863 images for human review and had an estimated total duration of 43.5 hours. It found 217 independent events compared with 378 manually. [1]

AI directly reduced the image set that people reviewed. The time comparison combines measured compute with estimated human review and does not hold detection quality constant. [1]

Human review within the hybrid workflow was estimated. The same reviewer assessed both methods. Ground truth combined their findings and could miss snakes overlooked by both. Results depend on species size, camera conditions, hardware, and threshold. Image recall and independent-event detection use different denominators. [1]

The image-review denominator is 1,485,924 images. Detection metrics use the subset remaining after humans, vehicles, and other animals were removed. Combined independent snake events number 447. [1]

The paper states that the study data are available on request. No public study dataset, analysis-code archive, or supplement is linked. Those unpublished materials were not inspected. [1]

The study does not report a monetary benefit or net-cost estimate. [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

Not reported [1]

University of Florida, South Florida Water Management District, and U.S. Geological Survey researchers evaluated the workflow. Dan Morris is acknowledged for technical advice and is not a study author. [1]

Three sites in Miami-Dade County, Florida, United States. Processing used UF HiPerGator. [1]

Not assessed

The study used 42 Reconyx HP2X cameras, UF HiPerGator, an Nvidia A100 GPU, MegaDetector 5A, Timelapse2, and expert human review. [1]

Specific cross-border transfers of study data or model outputs are not documented in the reviewed sources. [1]

Not assessed

The chosen threshold must retain the species and events that matter to the research question. Available computing and expert review must be included when comparing workflow time and cost. Images, labels, outputs, and timing records must be available to reproduce the comparison. [1]

Not assessed

Not assessed

The study and versioned model guide are public. The underlying study data are available on request. [1]

The study does not document a subsequent institutional or conservation-management decision. [1]

Not reported [1]

Not reported [1]

Not reported [1]

Two retained sources have stable identifiers beneath this record.

Miller et al., Methods, Tables 2-4, Discussion, and Data Availability. The versioned MegaDetector guide supplies model history and limitations. [1] [2]

The retained sources are in English. [1] [2]

The study authors are researchers at UF, SFWMD, and USGS, with no model-developer author listed. Dan Morris supplied acknowledged technical advice. The guide is developer documentation and does not independently confirm the benefit. [1] [2]

The study supplies the comparative results. The developer guide supplies model methods and does not independently confirm the benefit. [1] [2]

The original April 2025 study and February 27, 2025 model-guide version were inspected. Study data are available on request and were not inspected. [1] [2]

MegaDetector removed 94% of images from manual review in a Florida study, but the hybrid workflow detected fewer snake events overall. [1]

The original study supports the comparative result. The second source adds methods or reproducibility information. [1]

An original comparative field-data study supports reduced review work and documents detection losses. The hybrid human-time component is estimated, and the study does not independently measure financial or conservation outcomes. [1]

Actual hybrid review logs, complete workflow costs, subsequent adoption, and conservation decisions remain unresolved. Study data require a request to the authors. [1]

The documented research benefit for the participating organizations. [1]

Organizational depth 1: reduced review work. Societal depth is unassessed because a conservation outcome is not documented.

Researchers reviewed fewer images, and the study estimated a shorter workflow. Detection losses limit which monitoring tasks can use it. The study does not establish a conservation intervention or wildlife outcome.

The record distinguishes demonstrated analytical results from estimated or unmeasured downstream benefits. [1]

AIBD depth scale and the Trustible benefit taxonomy in the current local catalog, reviewed October 25, 2026.

AIBD reviewed this assessment on October 25, 2026.

Manual review detects more snake events overall. A hybrid workflow may suit particular large species, while a broader inventory may require more human review. [1]

Independently studied, with developer technical advice acknowledged and the timing estimate qualified.

Comparative research study. [1]

The other new conservation record concerns a separate evaluation that also uses MegaDetector. It is not a replication of this result.

Environment and conservation [1]

MegaDetector locates animals, people, and vehicles in camera images. This study tested whether filtering images could reduce the work required to monitor cryptic snakes. [1]

Images were collected from June 20 through July 18, 2022. The study was first published April 20, 2025. The AI-analysis completion date is not reported. [1]

Research evaluation using field-collected camera images. [1]

The record has a content fingerprint that changes when its source data change.

The study DOI is 10.1002/rse2.70009. The record also links the versioned MegaDetector guide. [1]

Not assessed

Not assessed

The record and its source assessment were added on October 25, 2026.