What SLU's report on drone surveys of moose says

Why the report was written

Drone surveys are already on sale, but nobody had tested whether they give a basis that holds up for management. The Swedish Environmental Protection Agency (Naturvårdsverket) therefore commissioned SLU in 2025 to analyse the method and propose a quality-assured way of surveying moose with drones (§1.1). The commission required, among other things, density, the adult sex ratio and calves per cow, each with a confidence interval, and an assessment of how the uncertainty depends on how much of the land is surveyed (§1.1).

Three areas, three different questions

The study was flown in the winter of 2026 in three areas (§2.2):

• Järnäshalvön moose management unit (ÄSO) in Västerbotten, 14 700 ha. A sample plot survey: 26 moose were observed in 43 sample plots totalling 2 327 ha, 16 % of the area (§3.1).
• Alster moose management unit (ÄSO) in Värmland, 22 400 ha. The whole area was flown manually in about 40 hours (§2.2.2), and 197 moose were observed (§3.2). Fifteen sample plots were also flown twice, once manually and once automatically, to compare the two approaches.
• An area outside Fredrika in Västerbotten, 3 400 ha, chosen because it had GPS-collared moose (§2.2.3). Two pilots flew the area in parallel, and twelve flights were aimed at the collared animals.

It is a sensible design. The complete counts in Alster and Fredrika give an answer key that samples can be simulated against, and the collared moose give an answer key for how many animals the drone actually sees.

The drone saw almost every moose, in the right weather

Of twelve flights aimed at GPS-collared moose, the animals were still in the square on eight, and all of them were found (§3.3). In Alster the automatic flight found the same moose as the manual one in all fifteen sample plots, and in Fredrika the agreement was 0.98 (§3.2, §3.3). The report sums up:

“The detection rate in this study was estimated at close to 1, based on both parallel flights over the same area and targeted flights in areas with GPS-collared moose.” (§4.1, our translation)

The condition is what matters. The results apply to cold, overcast winter weather, which gives high thermal contrast between animal and surroundings (§4.1). In sunnier conditions another study has reported a detection rate of around 0.8 (Mayer et al. 2024, cited in §4.1). The report therefore recommends that the detection rate be estimated under other conditions, for example by surveying about 25 % of the sample plots on repeated occasions (§4.1, §5).

How much of the land has to be flown

This is the report's most useful result, and it is worth reading carefully because two figures appear. The discussion says:

“Our results indicate that a coverage of about 20 % is sufficient in areas with moderate aggregation, while up to 50 % of the area may need to be surveyed in areas with strong aggregation to achieve sufficient precision in the estimates.” (§4.2, our translation)

The summary and the conclusions say about 20–30 % instead (§5). The difference says something in itself: how much is needed depends on how clustered the moose are, and the report says explicitly that it should be assessed for each area separately (§5).

The figures come from simulations on the complete counts. In Alster, where the moose were clustered, the report says 20 % of the area was enough for density, 30 % for the sex ratio and 40 % for calves per cow (§3.2; on the 30 % see further down). In Fredrika, where they were even more clustered, density needed 50–60 % (§3.3). Sex ratio and calves need more, because they rest on a subset of the animals (§4.4).

Precision counts as sufficient when the width of the confidence interval divided by the estimate is below 1.0, which roughly corresponds to a coefficient of variation below 0.3 (§2.3.3, §4.3). The report itself points out that this does not mean the precision is enough in every management situation, since different users have different requirements (§5).

The size of the sample plot matters less than the total coverage (§4.2). It is set mostly by the need to keep the drone within visual line of sight, and sample plots larger than about 50–60 ha are difficult in many forest settings without a permit to fly beyond visual line of sight (§4.2). The report suggests starting with a sample of, for example, 10 % of the area and working out the required coverage from how unevenly the moose turn out to be spread (§4.2).

The densities, and why most of the intervals are wide

Here are the report's density estimates, in moose per 1 000 ha with 95 % confidence intervals:

• Järnäshalvön, sample plots: 11.2 (3.8 to 20) (§3.1)
• Alster, whole area: 8.8 (3.0 to 14.6) (§3.2)
• Fredrika, whole area: 17.8 (9.5 to 26.1) (§3.3)
• Fredrika, sample plots: 7.8 (0.0 to 17) (§3.3)

It may seem contradictory that the intervals are so wide when the drone saw almost every animal. The report explains it plainly: the detection rate decides whether the animals in the sample plots are seen, but the precision is decided by how well the sample plots represent the whole area (§4.3). Few observed moose and an uneven spread give wide intervals, and Järnäshalvön and Fredrika were surveyed at 16 % and 23 % of the area respectively, at the low end of or below what the results later showed is needed (§4.3).

Do the densities agree with other methods? For two of the three areas there was something to compare with (§4.3). In Alster the earlier camera-trap estimate was around 9 moose per 1 000 ha. On Järnäshalvön camera traps gave about 6 and pellet counts 5 to 10, which the report judges reasonable given the low precision. None of the comparison figures is given with an interval in the report, so the comparison is a plausibility check and not a test.

The report also warns against reading the intervals of the complete counts too literally. They do not rest on variation between samples but on an approximation, and the true precision is probably higher (§4.3).

Sex and age can be determined, but it takes time

With a zoom camera, preferably from several angles, sex and age can be determined (§4.4). It needs experienced assessors, because the differences are small in winter when the bulls have no antlers, and the report recommends always taking photos so that the assessment can be reviewed afterwards (§4.4). On Järnäshalvön only 42 % of the observed animals could be classified (§3.1), and the difference from Alster is attributed mainly to the assessors having different thresholds for when an assessment counted as certain (§4.4).

What the drone companies do today

SLU worked with several operators and describes two main approaches (§3.4). One is sample plots laid out at random with 20–40 % coverage, flown manually, usually with a confidence interval and with sex ratio and calves per cow. The other is automated full-coverage flights where AI identifies the species, without sex and age determination. Most companies offer full-coverage surveys of smaller areas (§3.4).

The operators themselves ask for standardisation, certification of pilots and companies, and accepted survey designs, because today's results depend to a large degree on who is flying (§4.5).

What the report leaves open

The report is clear about its limits. Four knowledge gaps are identified (§5): more complete counts in different landscapes to see how the spread drives the coverage, detection rate in partly clear and clear weather, detection rate in different habitats and seasons, above all summer and broadleaf forest, and how accurate AI-based counting in thermal images is.

Two practical limits recur. The weather: during the study there were several weeks when no survey could be carried out, and cold below -15 °C can affect some drones (§5). The rules: the requirement to fly within visual line of sight limits the method, and the report recommends that Naturvårdsverket, together with the Swedish Transport Agency (Transportstyrelsen), look into flying beyond visual line of sight for wildlife surveys (§5).

And some things are not in the report at all. It gives no costs. It is not an approval of drone surveys as a basis for moose management: the conclusion is that drones have great potential as a complement to existing methods (the summary), and Naturvårdsverket has since said that more evidence is needed before the method can be approved (Svensk Jakt, 26 August 2026 (in Swedish)).

We have recomputed the report, and two values do not add up

We have built the report's calculations and get the same answer on 37 of 39 published values. The two that do not match affect none of the field estimates above, but we would rather state them than leave them unsaid.

• Calves per cow in Alster. The relative interval width is printed as 4.3 (§3.2). The estimate 0.48 and the interval 0 to 1.85 give 3.85.
• The coverage for sex ratio in Alster. The text says 30 % of the area was needed (§3.2). The table for sample plots of 50 ha, the size Figure 2 shows, meets the requirement already at 20 %: 0.37 with the interval 0.20 to 0.55, a relative width of 0.95 (Appendix 1). Even rounding the table in the unfavourable direction, the width stays below 1.0. The figure 30 % can be obtained if the threshold is read across several sample plot sizes together, and that is probably how it was arrived at.

From the report's foundation we have worked further on the points where it leaves the choice to whoever carries out the survey: the coverage is worked out for each area, the detection rate is measured instead of assumed, and the plot layout is tested before the price is set. Read how the method is built.

Read more about moose surveys

• The method behind the moose survey What we build on, and where we have taken it further.
• How many sample plots are needed (in Swedish) The arithmetic behind the number, and why a fixed rule misses.
• Compared with other ways to count moose (in Swedish) Helicopter, pellet counts and hunter observations, and when drones are the wrong choice.
• When in the year the survey is flown (in Swedish) What the weather and the light require, and why the season needs slack.
• How to read the survey report What a confidence interval means, and when a figure cannot be used.

Contact

hej@sparnum.com

Have you read the report differently, or found something we have missed? Get in touch. We would rather have it corrected than be right.