AI in Drones Statistics for Adoption Surveys and Image Benchmarks
By Avantika Kapoor · Published
WIRES reports five Australian wildlife hubs using drone footage for koala monitoring and AI analysis at Queensland University of Technology. This documents use in a named programme across three states. PwC’s UK survey and the HIT-UAV and VisDrone research datasets provide complementary evidence on commercial attitudes and evaluation samples. A global adoption rate would need a defined population of eligible operators.
Documented drone-imagery AI use, survey responses and research benchmarks
| Source and population | Published observation | Interpretation | Source |
|---|---|---|---|
| WildSeek · Australia, FY 2024–25 | 5 hubs across 3 states | Community hubs collected thermal drone footage for koala surveys. The university analysed footage using AI. | [4] |
| Opinium survey for PwC · UK, 2022 | 251 industry respondents | People working in sectors where drone services could be bought or supplied. | [1] |
| Same UK survey | 63% cited credible service providers | Response about a factor that would encourage use of commercial drone services. | [1] |
| HIT-UAV · 2023 dataset paper | 2,898 images from 43,470 frames | Thermal images selected for the dataset. The test split contains 579 images. | [2] |
| VisDrone · research team’s dataset | 10,209 static images and 288 video clips | The clips contain 261,908 frames. A clip and a static image are different observation units. | [3] |
Which percentage belongs in an adoption brief?
PwC’s 63% describes a stated condition for greater use. The survey also identifies 63 individuals who had worked with drones commercially or purchased drone services. Those two results happen to use the same digits, but one is a percentage and one is a count. The latter represents roughly one quarter of the 251-person sample. Neither question identifies a particular AI capability deployed on an aircraft. In the Australian programme the AI processing occurred at the university, so the five hubs document use of AI in a drone-imagery workflow while leaving onboard adoption unmeasured.
How large is an AI evaluation sample?
HIT-UAV retained one image per fifteen source frames on average. Its 2,029 training images, 290 validation images and 579 test images sum to the published 2,898. A result measured on the test split should carry the 579-image denominator. VisDrone publishes both still images and a video collection. Comparing counts across the projects requires specifying whether a row represents an image, a frame or a sequence.
Research questions
What would it take to report an AI-drone adoption rate?
Define the geography, period and eligible operators, then count respondents with a deployed, defined AI function. Separate product availability, trials and routine use. Report the sample size, question wording and any weighting used.
Can image counts establish which AI product performs better?
Image counts establish dataset size. Product performance requires results on the same evaluation task, classes, held-out sample and metric. Neither a larger public dataset nor a larger number of commercial survey respondents establishes a performance ranking.