147 DETECTING UNMARKED MOOSE WITH INFRARED SENSORS VIA AN UNOCCUPIED AERIAL SYSTEM AND CORRECTING FOR SIGHTABILITY Lily M. Hall1*, Franklin B. Sullivan2, Sophia A. Burke2, Michael W. Palace2,3, Henry Jones4, and Remington J. Moll1 1Department of Natural Resources and the Environment, University of New Hampshire, 56 College Road, Durham, NH 03824, USA; 2Earth System Research Center, University of New Hampshire, 8 College Rd, Durham NH 03824, USA; 3Department of Earth Science, University of New Hampshire, 56 College Road, Durham, NH 03824, USA; 4New Hampshire Fish & Game Department, 629B Main Street, Lancaster, NH 03584, USA *Corresponding author: Lily M. Hall, lilymhall32@gmail.com ABSTRACT: Accurate and precise estimates of moose (Alces alces) density are pivotal for under- standing population dynamics and informing management decisions. One promising tool for obtain- ing this information is unoccupied aerial systems (UASs). However, this technology still requires critical evaluation, especially regarding properly accounting for imperfect detection, i.e., the probabil- ity that moose are available but not detected and therefore uncounted. A recent review found that less than half of studies estimating moose density adequately accounted for imperfect detection, suggest- ing that many moose populations might be underestimated. Our objective was to create a sightability model for unmarked moose detected from a UAS equipped with a long-wave infrared sensor that included covariates expected to affect large-scale UAS moose surveys. We conducted 35 UAS flights in northern New Hampshire, USA, during January and February 2023 and completed sightability maneuvers over 59 moose detections to collect images at various relative observation angles. From a Bayesian logistic regression based on a naïve observer analysis, we found that greater conifer cover and sunnier conditions strongly reduced sightability of moose whereas ambient temperature had a weaker but also negative effect. Sightability was near 100% below a threshold of approximately 50% conifer cover, above which sightability declined rapidly which is similar to findings from previous work. This study provides the first successful quantification of sightability for a non-collared moose population, demonstrating a cost-effective approach for calibrating UAS sampling for additional locations and species while paving the way for future applications of this model to correct moose population sampling. ALCES VOL. 60: 147 – 165 (2024) Key Words: Alces alces, drone, imperfect detection, moose, sightability, thermal, ungulate, unoccupied aerial system (UAS) Estimating animal density and monitoring densities over time are pivotal for under- standing population dynamics and for mak- ing informed management decisions, especially for charismatic and harvested spe- cies impacted by emerging threats like cli- mate change (Dice 1938, Sala et al. 2000, Fryxell et al. 2014). Moose (Alces alces) are one such species given their intensive man- agement as a game animal, their susceptibil- ity to expected climate change effects, and their overall public interest (Boutin 1992, Rempel et al. 1997, Musante et al. 2010, Jones et al. 2019, Ruprecht et al. 2020, Moll et al. 2022). Beyond general ongoing work to monitor moose as a game species, recent mailto:lilymhall32@gmail.com MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 148 research effort has also focused on predict- ing potential range shifts of moose in response to climate change (Tape et al. 2016, Teitelbaum et al. 2021) and understanding moose population responses to parasitism by winter ticks (Dermacentor albipictus) (Jones et al. 2019, Ellingwood et al. 2020, DeBow et al. 2022). Topics like these would be fur- ther illuminated by obtaining improved spa- tially explicit density estimates for moose populations throughout their range (Michaud et al. 2014, Hinton et al. 2022). Such esti- mates have traditionally come through ground surveys of tracks or pellets or from aerial surveys using helicopters or fixed- winged aircrafts (e.g., Gasaway et al. 1986, Kretser et al. 2016). However, these meth- ods can be prohibitively time- and cost-in- tensive, and the predominant method – aerial surveys – is limited to optimal winter weather conditions that are diminishing due to climate warming (Brinkman et al. 2024). Considering logistical demands, these meth- ods can limit the spatial scale and temporal resolution of data that researchers and man- agers use to monitor moose (Rahman and Rahman 2021, Moll et al. 2022). Other more cost-effective monitoring techniques exist, including the use of public reports and har- vest data analyses. However, these approaches require careful and periodic cali- bration using independent datasets (Bontaites et al. 2000) and can be limited to areas with sufficient harvest (Solberg et al. 1999) or hunter observers (Boyce and Corrigan 2017) and thus often warrant supplementation through other monitoring methods. Recent technological developments in unoccupied aerial systems (UASs) hold promise for efficiently obtaining spatially explicit moose density estimates as well as increasing the temporal resolution of data while remaining at an accessible cost for many research entities (Anderson and Gaston 2013, Linchant et al. 2015, Rahman and Rahman 2021, Moll et al. 2022, Elmore et al. 2023). Given that UASs can fly at lower altitudes and slower speeds than piloted aircrafts, they can provide finer reso- lution data (Anderson and Gaston 2013) and can be safer and more adaptable in adverse weather conditions (Zmarz et al. 2018). In addition, UASs can be readily equipped with various sensors suited to detect wildlife (Iglay et al. 2024), including long-wave infrared sensors that enable detection of individual endotherms such as moose. Although UAS studies have examined free-ranging (McMahon et al. 2021b, Mayer et al. 2024) and captive ungulate populations at localized spatial scales (e.g., 0.13 km2, Zabel et al. 2023; 1.74 km2, Beaver et al. 2020), UAS surveys have generally not been conducted at larger spatial scales (Wang et al. 2019). An exception was a large-scale evaluation of kangaroo monitoring that con- cluded UAS surveys were an unsuitable replacement for their helicopter surveys, partially due to insufficient resolution of infrared imagery for species identification (Gentle et al. 2018). However, improve- ments in UAS, battery, and infrared sensor technologies have extended flight times and increased image resolution, making UASs a promising solution for monitoring species on broader spatial scales (Linchant et al. 2015, Cleguer et al. 2021), particularly for moose monitoring where species differentia- tion is more distinct. Nevertheless, difficul- ties related to environmental conditions such as snow presence, thermal contrast, and veg- etation obstruction must be navigated (Havens and Sharp 2015, Burke et al. 2019). A limitation of UAS sampling in the USA is the availability of launch locations with suf- ficient visibility to comply with the Federal Aviation Administration requirement to maintain visual line-of-sight (VLOS) with the system during flight (FAA 2024). The availability of such launch locations will ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 149 vary by land cover and is expected to be rar- est in tall, dense conifer-dominated forests. Therefore, although UASs are a promising new technology for monitoring moose and other wildlife species, their use requires fur- ther field validation. As for any wildlife survey attempting accuracy, a critical component of moose sur- veys is accounting for imperfect detection by estimating the probability of detecting a study animal that is present on a transect or sample plot (Caughley 1974, MacKenzie et al. 2002, Tyre et al. 2003, Cleguer et al. 2021). After biologists quantify this proba- bility, they can apply a sightability correc- tion factory (SCF; (Gasaway et al. 1986, Steinhorst and Samuel 1989) to their popula- tion estimate which is crucial for obtaining accurate abundance data (Guillera-Arroita et al. 2014) and obtaining reliable inference about species occurrence and wildlife-habi- tat relationships (Kéry et al. 2010). Although sightability corrections are particularly rele- vant to reducing bias due to undercounting, it is critical to also understand the potential of overcounting based on sampling design and animal movement (Schultz et al. 2024). Imperfect detection of moose has previously been addressed using double observers (Cumberland 2012, Kantar and Cumberland 2013, Oyster et al. 2018), calibrations with infrared technology (Bontaites et al. 2000), and quantifying detection probability via radio- or GPS-collared individual moose (Peters et al. 2014, Oyster et al. 2018, McMahon et al. 2021b). Steinhorst and Samuel (1989) pioneered an approach to correcting for imperfect detection by model- ing factors affecting detection using logistic regression that has come to be known as sightability modeling. However, many stud- ies estimating moose populations (including aerial, ground, and public surveys) have not accounted for imperfect detection. For example, a recent review of 89 moose monitoring studies found that only 36% explicitly accounted for imperfect detection and that accounting for detection bias had not improved over time (Moll et al. 2022). This lack of explicitly accounting for imper- fect detection is particularly concerning as even moose sightability models with similar covariates have been shown to produce very different corrected population estimates (e.g., 5.6x different, Harris et al. 2015). Furthermore, evaluations of emerging infra- red detection technology as an aerial survey- ing method (e.g., infrared detectors on UASs) have only recently been conducted (McMahon et al. 2021a, b; Delisle et al. 2023; Zabel et al. 2023) and are rare for moose (Elmore et al. 2023). Despite this, UASs equipped with infrared sensors can detect moose accurately in some conditions, with up to an 85% detection rate when flying in conditions that maximize thermal contrast such as cooler parts of the day and overcast skies (McMahon et al. 2021b). However, detection probability declines sharply with canopy closure which in the winter is often due to the presence of conifer trees (Peters et al. 2014, Oyster et al. 2018, McMahon et al. 2021b). Two studies have examined moose detection probability using infrared imagery with UASs (McMahon et al. 2021b, Mayer et al. 2024). Here, we build upon these initial assessments to create a sightability model of unmarked moose detected from a UAS equipped with an infrared sensor. We used naïve observer detections of known moose to examine a Bayesian logistic regression model that incorporated biotic and abiotic covariates that we hypothesized to influence moose sightability (Table 1). These covari- ates, including temperature, cloud cover, and conifer cover, would be expected to affect any large-scale UAS moose survey. Our results provide a model for use in future sampling surveys to estimate spatially MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 150 explicit moose densities that will be useful for informing moose management and guid- ing the application of UAS infrared technol- ogies to other systems. STUDY AREA We conducted UAS flights from 5 launch sites in northern New Hampshire, USA, in New Hampshire Fish and Game Wildlife Management Units A1, C1, and C2 (NHFG 2023). The study area was mostly forested with low levels of anthropogenic develop- ment; land cover consisted of approximately 91% forest (16% conifer, 34% deciduous, and 41% mixed), 3% developed, 3% shrub/ scrub land, and 2% water (Dewitz 2020). The predominant forest type was northern hardwood-conifer forest consisting mainly of maple (Acer spp.), birch (Betula spp.), and American beech (Fagus grandiflora) with lowland and high elevation spruce-fir forests of primarily red spruce (Picea rubens) with intermixed balsam fir (Abies balsamea; DeGraaf and Yamasaki 2001, NHFG 2015, Jones et al. 2019). Launch location elevations ranged from 400 to 700 m. Snowfall during the study period was 95 cm with a mean temperature of -5°C while Table 1. Covariate hypothesis, name of covariate, description of data used in modeling to operationalize the covariate hypothesis, the predicted positive (+) or negative (-) effect of the covariate on moose sightability, and relevant references for hypotheses evaluated for predicting moose sightability from infrared images collected via an unoccupied aerial system for moose in New Hampshire, USA, winter 2023. A summary of the data is included in the covariate column. Covariate hypothesis Covariate Data description Predicted effect References The thick canopy of conifer trees obscure infrared sensing of objects below Percentage of conifer cover (min = 24, max = 85, mean = 61, SD = 18) The percentage of the RGB image pixels identified as conifer via GLI - Kissell and Nimmo 2011, Peters et al. 2014, Oyster et al. 2018, McMahon et al. 2021b, Hinton et al. 2022 Thermal contrast between moose and the environment decreases as temperature increases reducing visibility Ambient temperature (min = -8.9, max = 2.2, mean = -1.7, SD = 3.1) The ambient temperature in °C during image collection - Street et al. 2015, Hinton et al. 2022, Zabel et al. 2023 Ambient temperature and temperature of objects in the environment increase with time since sunrise, decreasing thermal contrast Time of day (min = 6.0, max = 16.0, mean = 11.0, SD = 2.6) The hour of the day rounded down to the whole hour - Zabel et al. 2023 Cloud cover blocks solar radiation, increasing thermal contrast Cloud cover (n = 707 overcast; n = 158 partly sunny) Categorical degree of cloud cover at the start of image collection + Millette et al. 2011, McMahon et al. 2021b, Mayer et al. 2024 Temperatures decrease into February and increase thermal contrast Day of year (min = 16, max = 46, mean = 27, SD = 6.3) Julian day of image collection + Quayle et al. 2001, Zabel et al. 2023 Fewer observation angles reduce the chance of the moose being detected due to environmental obstruction Unequal (n = 420 unequal; n = 445 equal) Whether or not the sequence had moose not in the viewshed for all 3 images - Coates et al. 2019 ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 151 the 1991 - 2020 average annual temperature was 6°C (NOAA 2024). The area was con- sidered high quality moose habitat with the 2022 estimated moose densities ranging from 0.6 - 3.5 moose/km2 across the region (NHFG 2023). Landownership included state owned parks, privately owned land, and commer- cially logged forests. We secured written per- mission from all landowners to collect data from a UAS prior to flying. To maximize the area covered per flight, we focused sampling in Jericho Mountain State Park (latitude: 44.47, longitude: -71.26) which provided optimal launch locations due to a clear view over the canopy from across a large lake (Jericho Lake) and also enabled us to adhere to the Federal Aviation Administration regu- lation of maintaining a visual line-of-sight of the UAS during flights. METHODS UAS Data Collection During January and February 2023, we con- ducted 35 UAS flights using a DJI Matrice 300 RTK quadcopter (Shenzhen DJI Sciences and Technologies Ltd., Nanshan District, Shenzhen, China) equipped with a DJI Zenmuse H20T sensor perpendicular to the ground recording color (RGB; 4056 x 3040 pixels, 82.9° DFOV) and long-wave infrared thermal images (640 x 512 pixels, 40.6° DFOV). For each flight, we confirmed consistent snow cover but did not have a required minimum depth and recorded ambi- ent temperature (°C), relative humidity, and windspeed (mph) at the beginning of the sightability maneuver with a handheld Kestrel 5500 weather station (Kestrel Instruments, Boothwyn, Pennsylvania, USA) and noted time of day and sky condi- tion as potential categories of overcast, partly sunny, sunny, or dark. We started flights by following pre-pro- grammed, terrain-following lawnmower pattern flight paths created in UgCS (v. 4.8.728, Smart Projects Holding, Ltd., Valletta, Malta) and imported to DJI Pilot 2 (Shenzhen DJI Sciences and Technologies Ltd., Nanshan District, Shenzhen, China) as KML files with the objective of detecting moose for further data collection as opposed to abundance sampling. We set photogram- metry mission parameters targeting 10 cm resolution for thermal images, 20% side overlap, and 33% forward overlap to balance the goals of searching as much area as possi- ble while having sufficient detectability for moose. This resulted in flying at ~ 100 m altitude at 8 m/s while recording images every 2 seconds. After we detected a moose, we manually flew a set pattern known as a sightability maneuver to mimic intensive flights in previous studies (Gasaway et al. 1986, Poole et al. 1999, Bontaites et al. 2000). These flights assumed a 100% detec- tion rate by increasing sampling effort (e.g., flying lower or slower) to generate an abun- dance estimate to which the corresponding sampling flight (e.g., flying faster or at higher altitudes) would be compared. Our sightability maneuvers made repeated flights at ~ 100 m altitude over confirmed moose detections (i.e., true positives; n = 59) that would later be used to estimate moose sightability in a naïve observer analysis. During the sightability maneuvers, we cap- tured simultaneous infrared and RGB images with the sensor downward-facing (i.e., nadir) and manipulated the flight patterns to posi- tion moose detections across all areas of the sensor viewsheds (i.e., in the left, right, top, bottom, center, and all corners of a square viewshed). We also flew true negative sightability maneuvers over the locations where true positive sightability maneuvers were conducted at least 6 days after the moose left. These true negative maneuvers MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 152 enabled us to assess detection rates across identical vegetation conditions, thereby con- trolling for unmeasured factors related to landscape features. Finally, we estimated the percentage of conifer cover for all images captured during flights by: (1) retrieving and clipping a cor- responding wide view RGB image to the footprint of the thermal image; (2) calculat- ing the median Green Leaf Index (GLI, Eq. 1, e.g. Louhaichi et al. 2001) of a 25-pixel window around each pixel; and (3) calculat- ing the proportion of all pixels in the image with a GLI ≤ 0.75. GLI = ((GDN – RDN) + (GDN – BDN)) / ((2*GDN) + RDN + BDN) (Eq. 1) where DN is the digital number of each band (i.e., Blue, Green, and Red) preceded by an identifier of the band (B, G, and R respectively). GLI is a useful index for detecting chlorophyll content of leaves and vegetation because chlorophyll absorbs blue and red wavelengths and reflects green wavelengths (Louhaichi et al. 2001, Macfarlane and Ogden 2012, Bush et al. 2020). This GLI threshold approach has been previously applied to winter optical imagery as a simple method to differentiate conifer and deciduous canopy vegetation (Sullivan et al. 2023). Naïve Observer Analysis We randomly selected 300 infrared images with moose as true positives and 300 with- out moose as true negatives (i.e., controls that lacked moose) for the naïve observer analysis. The true positive images were in three-image sequences to create 100 sequences, with each sequence displaying 1 focal moose being moved down each third of the image viewshed to replicate the changes in perspectives that occur for an observer during forward flight. Similarly, we prepared 100 true negative image sequences by choosing a focal ground location instead of a focal moose to display. The sequences were distributed so each vertical third of the viewshed was represented evenly. Percentage of conifer cover was also distributed evenly across the possible 0 – 100% range by lever- aging multiple sequences from some trials, although values in the extreme regions of the percentage of conifer cover (below 23% and above 85% conifer cover) were not surveyed in the field and thus not available. We ano- nymized and randomized the thermal and RGB images from true positive and true negative sightability maneuvers using pack- age magick (Ooms 2024) in R (version 4.2.1; R Core Team 2024). We then provided the infrared images to naïve observers using the program timelapse (https://timelapse.ucal- gary.ca, version 2.3.0.6; Greenberg et al. 2019). We trained naïve observers (n = 11) to identify moose in infrared image sets via a 45-minute video (see Supplemental Material) containing explanations on image features useful for identifying moose (i.e., shape, size, contrast from background, tracks). We overlaid a light gray 3 x 3 rectangle grid over the infrared images and compared the true moose presence to the naïve observer recorded moose presence in each of the 9 cells. This defined a smaller area to help determine if a naïve observer identified a moose in the image as opposed to a non- moose feature. These observers had no prior knowledge of image locations or whether an image contained a moose (McMahon et al. 2014). Naïve observers had the option to examine both the infrared images and the corresponding RGB images during training and analysis (Fig. 1). We permitted naïve observers to examine RGB images at their discretion because we determined it helpful https://timelapse.ucalgary.ca https://timelapse.ucalgary.ca ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 153 in improving the accuracy of moose identifi- cation, particularly when a moose bed was present that often showed a similar signature in the infrared image to a moose. We included multiple naïve observers so that we could account for observer-level differences in detection rates (Attard et al. 2024). Prior to analyzing the naïve observer data to estimate moose sightability by cor- recting for false negatives, we first deter- mined if naïve observers recorded any false positives (i.e., identified moose that were not present) to confirm that such detections did not occur at a rate warranting inclusion in the model. False positives occurred < 1% of the time and this rate was too low to for- mally include in the sightability model. Additionally, in 10 out of the 875 sequences (1%), naïve observers recorded more moose than were present (e.g., a naïve observer recorded 3 moose when only 2 were pres- ent). We excluded these false positives in the sightability model as well, but we dis- cuss these false positive rates when Fig. 1. Corresponding long-wave infrared (A & C) and RGB (B & D) images captured while conducting UAS sightability maneuvers over moose (located within the yellow circles) in northern New Hampshire, USA, winter 2023. Each pair of images was captured simultaneously at the same location over either a high (A & B) or low conifer area (C & D), illustrating variation in moose sightability. The infrared images include a light gray 3 x 3 rectangle grid used in the naïve observer analysis to aid in image tracking. Image C also shows moose tracks in snow which was one sign naïve observers were trained to use as an aid in moose detection. MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 154 providing recommendations for future UAS work. After each naïve observer processed the 600 moose images (200 sequences of 3 images), we realized that during our sightability maneuvers (which we con- ducted following a zig-zag pattern from top- to-bottom, left-to-right), moose had more time to move between vertically sequential images than there would be in a traditional transect sampling approach. To maximize the applicability of our sightability model, we removed sequences in which moose moved more than 30% of the image length throughout the sequence (n = 5). We also removed images collected in the dark (n = 8) because flights pre- and post-sunrise were logistically difficult but did not seem to increase the thermal signature of moose. We therefore conducted the final sightability analysis on 261 unique true positive images (87 unique sequences). Sightability Model Description We analyzed the proportion of moose that naïve observers correctly identified (i.e., sightability) in true positive sequences using a binomial logistic regression model. We included covariates that we hypothesized would affect sightability in the linear predic- tor of the model (Table 1), as follows: mooseDeti ~ Binomial(pi, Ni) logit(pi) = βNO[i] + β.ConNO[i] * Coni + β.TimeNO[i] * Timei + β.Day NO[i] * Dayi + β.TempNO[i] * Tempi + β.UnequalNO[i] * Unequali + β.CloudNO[i] * Cloudi (Eq. 2) where mooseDeti was the number of total moose that the naïve observer successfully detected in the ith image sequence given pi as the probability of success (i.e., correctly identifying a moose that was truly present, also referred to as sightability) and a true number of moose, Ni. In this model formula- tion, each true positive sequence (n = 87 per naïve observer) served as a data point com- posed of Ni independent trials of moose detection, each with binary responses (detected or not). The intercept and each covariate effect (i.e., the βs) were all indexed by naïve observer (NO[i]) because we mod- eled these parameters as random effects to account for potential pseudo-replication due to different naïve observers processing the same image sequences. Accordingly, we made inference on parameters averaged across all naïve observers (i.e., the posterior of the random effect; see below). Coni, Timei, Dayi, Tempi, Unequali, and Cloudi were covariate values for the ith image sequence for percentage of conifer cover, hour of the day, day of the year, ambient temperature (°C), categorical classification of whether the number of moose were equal across all 3 images, and cloud cover (overcast or partly sunny, as dark images were removed and no moose were encountered in sunny condi- tions), respectively. We included the unequal covariate to account for sequences with an unequal number of moose per image (i.e., 1 moose might have only been present in 1 or 2 images in the sequence of 3 because a sec- ond moose was the one we targeted to be in each third of the viewshed). This situation would not arise with typical field data where the sampling transect would continue, so this covariate controlled for the nuisance effect of reduced number of observation angles for some moose to avoid bias in the sightability model. We used the package car (Fox et al. 2024) to check for collinearity between covariates by first identifying covariates with an adjusted generalized standard error inflation factor (aGSIF) > 1.7, which is com- parable to a variable inflation factor (VIF) > 3 (Fox and Monette 1992, Zuur et al. 2010). We also initially considered a nonlin- ear effect (Heit et al. 2024) of time of day by including a Time2 covariate to test the ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 155 hypothesis that time had a quadratic rela- tionship with moose sightability due to changing sunlight and temperatures through- out the day. However, Time and Time2 had a GSIF > 1.7, thereby introducing collinearity, and models including both terms did not converge in the Bayesian analysis described below. Prior to omitting Time2 from further consideration, we examined a frequentist version of the global model with and without the term and found no support for retaining Time2 based upon AIC (ΔAIC < 2; Arnold 2010). We therefore removed Time2 from further analysis, which resulted in all remain- ing covariate aGSIFs < 1.7. We analyzed the model using Markov Chain Monte Carlo (MCMC) simulations in a Bayesian framework. We used R to run the model in JAGS language with R2jags (Su and Yajima 2024). For the final model eval- uation, we used 3 MCMC chains of 10,000 iterations each with a burn-in of 1,000 and a thinning rate of 1 (Link and Eaton 2012). We used diffuse logistic distribution priors for all covariate parameters and scaled the covariates to facilitate comparison of effect sizes across covariates (Northrup and Gerber 2018). We confirmed model convergence by visually inspecting traceplots and ensuring that R-hat statistics were < 1.1 (Gelman and Hill 2007). We assessed model fit using Bayesian p-values using standard techniques (Kéry and Royle 2015). To do so, we com- pared the Pearson residuals from field-col- lected data to those calculated from model-predicted data. We considered model fit acceptable if the Bayesian p-value was 0.05 ≤ p ≤ 0.95 and excellent if the value was near 0.5, which would indicate that the model could faithfully reproduce the field data (Kéry and Royle 2015). We interpreted models by evaluating covariate effects (β values in Eq. 2) for significance, defined as whether the 95% credible interval contained zero or not (Kéry and Royle 2015). RESULTS Naïve observers correctly identified 1,074 of the 1,633 moose (66%) across 865 true pos- itive sequences of long-wave infrared images. A Bayesian p-value of 0.46 indi- cated that the model described above pro- vided an excellent fit to these data. We found that percentage of conifer cover had strong and significant negative effects on moose sightability (Table 2; standardized β poste- rior mean = -1.51). Sightability was near 100% below a threshold of approximately 50% conifer cover, above which sightability declined rapidly (Table 2; Fig. 2a). We also found that ambient temperature had a signif- icant but relatively modest negative effect Table 2. Estimates of the effects of scaled covariates on moose sightability from a binomial logistic regression fit to unoccupied aerial system data, winter 2023, New Hampshire, USA. Intercepts represent the sightability of moose on the logit scale with overcast skies and an equal number of moose in all image sequences. An asterisk (*) indicates significant terms (95% credible interval (CI) does not include zero). Posterior Mean Lower 95% CI Upper 95% CI Intercept 1.43* 0.71 2.15 Percentage of conifer cover -1.51* -1.96 -1.15 Time of day 0.16 -0.08 0.42 Day of year -0.27* -0.56 -0.01 Ambient temperature -0.24* -0.41 -0.07 Unequal presence of moose within the image sequence -0.36 -0.81 0.04 Cloud cover -1.81* -2.28 -1.37 MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 156 (standardized β posterior mean = -0.24) on sightability, which declined approximately 10-15% as temperature increased from -10° to 4° C (Table 2; Fig. 2b). Cloud cover had a strong effect on sightability (Table 2; Fig. 2c), with mean sightability during overcast conditions (80.8%) nearly double that during partly sunny conditions (40.6%). Day-of- year had a negative effect on sightability, but the relationship was modest (standardized β posterior mean = -0.27) and the 95% credi- ble interval very nearly overlapped zero (upper bound of -0.01). The model did not detect an effect of time-of-day or unequal presence of moose (Table 2). For the original 198 true positive and true negative sequences given to all observers, naïve observers accessed the RGB images in 11% of the sequences (232 of 2178 sequences). The observers changed their moose identifica- tion decision based on information from the RGB images in 2% of the sequences (52 of 2178 sequences). DISCUSSION We created a sightability model for unmarked moose from a naïve observer analysis of long-wave infrared images captured via a UAS. Because our flights were conducted during winter with snow cover and decidu- ous leaf-off, moose were generally easy to detect as large, warm bodied individuals in infrared imagery except when in dense coni- fer cover. Specifically, according to a Bayesian logistic regression, we found that a greater percentage of conifer cover and sun- nier sky conditions strongly reduced sightability of moose whereas ambient tem- perature had a weaker negative effect. Although most of these influences were expected (e.g., negative effects of conifer cover; McMahon et al. 2021b), determining location specific sightability models is important for accuracy as even models with similar covariates can greatly affect the cor- rected population count based on their respective coefficients (Harris et al. 2015). Fig. 2. Model-predicted sightability (probability of correctly detecting a moose) as a function of A) percentage of conifer cover, B) ambient temperature, and C) cloud cover based upon a binomial logistic regression model fit to data collected by an unoccupied aerial system during winter 2023, New Hampshire, USA. Solid black line represents mean predicted sightability, light grey lines depict uncertainty, and dashed lines depict extrapolated model predictions beyond the range of field data values. ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 157 Because the successful estimation of sightability for non-collared moose popula- tions had not been completed yet (McMahon et al. 2021b, Mayer et al. 2024), our results are an encouraging demonstration of a cost-effective approach for calibrating UAS sampling for additional locations and species The percentage of conifer cover, cloud cover, and ambient temperature were signif- icant covariates in the logistic regression model. Conifer cover was influential and negatively correlated with moose detection. This was expected due to visual obstruction (Cilulko et al. 2013, Burke et al. 2019) which has been the predominant factor influencing bias of UAS imagery of wildlife (Cleguer et al. 2021, Elmore et al. 2023) and previous findings of canopy cover effects (Chrétien et al. 2016, Doull et al. 2021, McMahon et al. 2021b, Mayer et al. 2024). The influ- ence of the percentage of conifer cover quantified as a relatively small area (40 m x 60 m viewing area of each image) highlights the importance of examining fine-scale covariates when accounting for sightability of wildlife during larger scale population sampling efforts. These fine-scale effects are likely because the conifer trees immediately surrounding the moose determine whether the moose is visible or not from above. However, future studies applying a sightabil- ity model to population sampling could build on our results by examining the scale of effect (Holland et al. 2004, Levin 1992, Wiens 1989) of percentage of conifer cover especially when considering varying levels of forest patchiness. Our observation of conifer cover effects is similar to, but slightly weaker than, that reported by McMahon et al. (2021b) for UAS field trials conducted with collared moose in Minnesota, USA. Additionally, overcast days have been shown to significantly increase moose detection with infrared sensors compared to sunny days (McMahon et al. 2021b, Mayer et al. 2024). This effect is likely caused by cloud cover reducing thermal loading which in turn increases the thermal contrast between the moose and its environment (Garner et al. 1995, Burke et al. 2019). In line with this expectation, we found that moose sightabil- ity was significantly greater on overcast days than on partially overcast days. Although we were unable to include sunny conditions in our model due to a lack of opportunities during data collection, we would expect sightability to be lower during full sun con- ditions as well. Finally, the model suggested a weak negative effect of ambient tempera- ture which we had predicted. This effect could result when warmer ambient tempera- tures decrease the thermal contrast between warm moose and their environment (Garner et al. 1995). Warmer temperatures also heat up other objects such as tree trunks or rocks that could provide a bright thermal signature similar to a moose in the infrared image (Garner et al. 1995, Doull et al. 2021). Interestingly, previous studies did not include ambient temperature in their best model; however, their UAS flights were conducted in generally warmer temperatures (e.g., mean [min, max]: 14 [-1, 27 °C] McMahon et al. 2021b; 14 [9, 23 °C] Mayer et al. 2024) than ours (-2 [-9, 2 °C]). This suggests that the influence of ambient tem- perature may plateau at higher temperatures. We observed more thermal noise, especially around the leafless deciduous canopies during our high temperature flights, which would decrease sightability of animals on the ground. Day-of-year, time-of-day, and unequal presence of moose had marginal or insignif- icant effects on sightability. Day-of-year had a significant negative effect on sightability according to our model, but the effect was highly variable and uncertain (Table 2). We had predicted that day-of-year would have a MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 158 negative effect because ambient tempera- tures in our study area of New Hampshire typically decrease through January with a slight increase in temperatures during the first weeks of February when we were sam- pling (NOAA: National Oceanic and Atmospheric Administration 1991). Additionally, environmental changes cor- related with date could influence animal behavior that would influence sightability (Cilulko et al. 2013). For example, moose have been shown to preferentially travel beneath conifer trees later in winter after accumulated snow in open areas impede movement (Balsom et al. 1996). We expected this behavioral change to reduce sightability by increasing the amount of time moose are blocked by conifer cover. We did not collect snow depth data, but anecdotally, moose sightings and tracks remained prevalent in open areas throughout the study period sug- gesting that snow depth had not reached a threshold to noticeably shift moose habitat use to more conifer areas. Similarly, our results did not support our prediction that time-of-day (on an hourly scale) would have a negative relationship with sightability due to generally increasing temperatures after sunrise which would decrease thermal con- trast. We had originally hypothesized that as temperature peaked during midday, sightability would also display this trend rel- ative to the time of day; however, we found no support for a quadratic relationship. This lack of a trend may be because our latest sample was at 16:00 and did not adequately capture the temperature decline towards the end of the day. Time-of-day might also be important in sunnier conditions because there would be greater thermal loading on vegetation, potentially decreasing thermal contrast between vegetation and moose. Our field conditions were typically cloudy or partly sunny, so we suggest researchers in sunnier conditions still consider the possibility of time-of-day effects. We con- clude that ambient temperature was more influential for sightability than day-of-year or time-of-day, although future work could provide greater clarity into these variables given their potential to covary. The predom- inant effect of ambient temperature would be expected given that it is more directly related to thermal contrast than time-of-day or day-of-year. Indeed, these latter 2 vari- ables are likely coarser proxies for changes in ambient temperature. We therefore sug- gest that including temperature in sightabil- ity models for thermal surveys is likely sufficient to account for these dynamics unless animal behavioral changes are expected to be strong across different times of day or year. Finally, our variable unequal was included in the model as a nuisance variable to account for the potential effect of having some moose visible in fewer than all of the 3-image sequences. This would not occur in typical sampling, where all image sequences would be viewed sequentially. Overall, these effects (percentage of conifer cover introducing visual obstruction and cloud cover influencing thermal contrast) corroborate those previously found in moose sightabilty models informed by UAS ther- mal cameras (McMahon et al. 2021b, Mayer et al. 2024). With the rapid development of artificial intelligence software, processing images with computer vision algorithms could reduce the time required for manual review (LeCun et al. 2015, Longmore et al. 2017, Lamba et al. 2019). Although not as appro- priate for species found in large aggregations (Attard et al. 2024), deep learning software that could classify images into presence or absence could still prove helpful for reduc- ing processing time of images moose (which are typically solitary). Regardless, these upcoming processing approaches will need to have human observers validate their ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 159 animal classifications to determine accuracy and precision. Currently, many UAS researchers determining sightability of wild- life from infrared sensors use a naïve observ- ers to process images prepared by a knowledgeable observer (i.e., someone who knows which files contain the target species) to quantify influential sightability factors or compare performance to traditional methods (Spaan et al. 2019, Doull et al. 2021). Most studies have found that infrared sensors or UAS surveys improve accuracy or process- ing time compared to traditional methods such as visual/RGB identification, ground surveys, or piloted aircraft surveys (Hodgson et al. 2016, Seymour et al. 2017, Elmore et al. 2023). To aid in the generalization of our model for potential future applications, we accounted for differences in naïve observer behaviors and abilities by averag- ing our 11 observers together as a random intercept. We also trained naïve observers in identifying moose in infrared images with the a single 45-minute training video. This replicable training approach may be useful in future applications. Our approach to quantifying sightability could be applied to population assessments either by surveying a small area and extrap- olating, or by surveying a larger area (Linchant et al. 2015, Spaan et al. 2019, Elmore et al. 2023). If live counts of moose (i.e., those done in real-time via video feed in the field by UAS pilots) were found to be comparable to the sightability-corrected naïve observer counts, then it would save post-processing time of future surveys. It is also possible that video data (e.g. Baldwin et al. 2023) or live counts would improve detection compared to only viewing 3 images. To make the process more compara- ble to visual observers from helicopters or fixed wing aircraft, it may also be worth the time and battery for UAS operators to pause for to confirm species identification. An additional consideration for improving this approach is reducing the false positive rate, even though it was very low in this study (< 1%). Previous work with identifying humans in infrared imagery found that naïve observers had a greater rate of false nega- tives but fewer false positives than an auto- mated machine learning analysis (Doull et al. 2021). Here, false positives were most commonly an empty moose bed that regis- tered an infrared signal resembling a moose. However, moose beds that would still appear white from snow were readily distinguish- able from dark brown moose in the RGB image, suggesting that false positives can be minimized by confirming moose presence in the paired RGB image if collected during adequate daylight. Similarly, fusing infrared with RGB images has been shown to improve animal identification by deep learning clas- sification networks compared to either data type alone (Krishnan et al. 2023). Both find- ings highlight the benefits of utilizing a dual sensor platform. Our study builds upon the growing body of literature showing the high potential of UASs to augment animal monitoring and population ecology studies. Moose are an excellent candidate for additional UAS mon- itoring because of their large body size and relative consistency of detection via infrared sensors. However, care must be taken to account for conifer and cloud cover effects given their strong influence on sightability. Sightability models such as the one pre- sented here can be readily applied to correct future sampling, which can enable UASs to more efficiently sample larger landscapes. ACKNOWLEDGEMENTS We thank M. Diaz, J. Merriman, J. Vincent, R. Andruk, F. Shinost, W. Chrisman, M. Poisson, A. Butler, S. Richard, C. Dawson, E. McCay, S. Young, and M. Legagneur for data analysis MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 160 support. We also thank J. DeBow for field assistance and landowners for permitting us to conduct this research and two reviewers for their feedback that helped improve this manu- script. This work was supported by the University of New Hampshire Agricultural Experiment Station under the USDA National Institute of Food and Agriculture McIntire- Stennis Project 7003422. This is Scientific Contribution number 3071. LITERATURE CITED Anderson, K., and K. J. GAston. 2013. Lightweight unmanned aerial vehicles will revolutionize spatial ecology. Frontiers in Ecology and the Environment 11:138–146. Arnold, T. W. 2010. Uninformative parame- ters and model selection using Akaike’s information criterion. The Journal of Wildlife Management 74: 1175–1178. AttArd, M. r. G., r. A. PhilliPs, e. Bowler, P. J. ClArke, h. CuBAynes, d. w. Johnston, and P. T. Fretwell. 2024. Review of satellite remote sensing and unoccupied aircraft systems for count- ing wildlife on land. Remote Sensing 16: 627. BAldwin, r. w., J. t. BeAver, M. MessinGer, J. MudAy, M. windsor, G. d. lArsen, M. r. silMAn, and T. M. Anderson. 2023. Camera trap methods and drone thermal surveillance provide reliable, comparable density estimates of large, free-ranging ungulates. Animals 13: 1884. BAlsoM, s., w. B. BAllArd, and H. A. whitlAw. 1996. Mature coniferous for- est as critical moose habitat. Alces 32: 131–140. BeAver, J. t., r. w. BAldwin, M. MessinGer, C. h. newBolt, s. s. ditChkoFF, and M. R. silMAn. 2020. Evaluating the use of drones equipped with thermal sensors as an effective method for estimating wild- life. Wildlife Society Bulletin 44: 434–443. BontAites, k. M., k. A. GustAFson, and R. MAkin. 2000. A Gasaway-type moose survey in New Hampshire using infrared thermal imagery: preliminary results. Alces 36: 69–75. Boutin, S. 1992. Predation and moose popu- lation dynamics: a critique. The Journal of Wildlife Management 56: 116–127. BoyCe, M. s., and R. CorriGAn. 2017. Moose survey app for population monitoring. Wildlife Society Bulletin 41: 125–128. BrinkMAn, t. J., k. A. kellie, A. k. reinkinG, G. e. liston, and N. T. BoelMAn. 2024. Changing snow conditions are challenging moose (Alces alces) surveys in Alaska. Wildlife Society Bulletin 48: e1555. Burke, C., M. rAshMAn, s. wiCh, A. syMons, C. theron, and S. lonGMore. 2019. Optimizing observing strategies for monitoring animals using drone- mounted thermal infrared cameras. International Journal of Remote Sensing 40: 439–467. Bush, e. r., e. t. A. MitChArd, t. s. F. silvA, e. diMoto, P. diMBondA, l. MAkAGA, and K. ABernethy. 2020. Monitoring mega-crown leaf turnover from space. Remote Sensing 12: 429. CAuGhley, G. 1974. Bias in aerial survey. The Journal of Wildlife Management 38: 921–933. Chrétien, l.-P., J. théAu, and P. MénArd. 2016. Visible and thermal infrared remote sensing for the detection of white-tailed deer using an unmanned aerial system. Wildlife Society Bulletin 40: 181–191. Cilulko, J., P. JAniszewski, M. BoGdAszewski, and E. szCzyGielskA. 2013. Infrared thermal imaging in studies of wild ani- mals. European Journal of Wildlife Research 59: 17–23. CleGuer, C., n. kelly, J. tyne, M. wieser, d. Peel, and A. hodGson. 2021. A novel method for using small unoccupied aerial vehicles to survey wildlife species and model their density distribution. Frontiers in Marine Science 8: 640338. ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 161 CoAtes, P. s., G. t. wAnn, G. l. Gillette, M. A. riCCA, B. G. ProChAzkA, J. P. severson, k. M. Andrle, s. P. esPinosA, M. l. CAsAzzA, and D. J. delehAnty. 2019. Estimating sightability of greater sage-grouse at leks using an aerial infra- red system and N-mixture models. Wildlife Biology 2019: wlb.00552. CuMBerlAnd, R. E. 2012. Potvin double-count aerial surveys in New Brunswick: Are results reliable for moose? Alces 48: 67–77. deBow, J., J. Blouin, e. rosenBlAtt, C. AlexAnder, n. Fortin, k. Gieder, J. MurdoCh, and T. donovAn. 2022. Birth rates and neonate survival in a parasite rich moose population in Vermont, USA. Alces 58: 51–73. deGrAAF, R. M., and M. yAMAsAki. 2001. New England wildlife: habitat, natural history, and distribution. University Press of New England, Hanover, NH. delisle, z. J., P. G. MCGovern, B. G. dillMAn, and R. K. swihArt. 2023. Imperfect detection and wildlife density estimation using aerial surveys with infrared and visible sensors. Remote Sensing in Ecology and Conservation 9: 222–234. dewitz, J. 2020. National land cover data- base (NLCD) 2016 products (ver. 3.0, November 2023). U.S. Geological Survey. diCe, L. R. 1938. Some census methods for mammals. The Journal of Wildlife Management 2: 119–130. doull, k. e., C. ChAlMers, P. FerGus, s. lonGMore, A. k. Piel, and S. A. wiCh. 2021. An evaluation of the factors affect- ing ‘poacher’ detection with drones and the efficacy of machine-learning for detection. Sensors 21: 4074. ellinGwood, d., P. Pekins, h. Jones, and A. MusAnte. 2020. Evaluating moose Alces alces population response to infestation level of winter ticks Dermacentor albi- pictus. Wildlife Biology 2020: 1–7. elMore, J. A., e. A. sChultz, l. r. Jones, k. o. evAns, s. sAMiAPPAn, M. B. PFeiFFer, B. F. BlACkwell, and R. B. iGlAy. 2023. Evidence on the efficacy of small unoc- cupied aircraft systems (UAS) as a sur- vey tool for North American terrestrial, vertebrate animals: a systematic map. Environmental Evidence 12: 3. FAA: Federal Aviation Administration. 2024. 14 CFR part 107 -- small unmanned aircraft systems. https:// www.ecfr.gov/current/title-14/part-107. Accessed 15 Nov 2024. Fox, J., and G. Monette. 1992. Generalized collinearity diagnostics. Journal of the American Statistical Association 87: 178–183. ____, J., s. weisBerG, B. PriCe, d. Adler, d. BAtes, G. BAud-Bovy, B. Bolker, s. ellison, d. Firth, M. Friendly, G. GorJAnC, s. GrAves, r. heiBerGer, P. krivitsky, r. lABoissiere, M. MAeChler, G. Monette, d. MurdoCh, h. nilsson, d. oGle, B. riPley, t. short, w. venABles, s. wAlker, d. winseMius, A. zeileis, and R-Core. 2024. car: Companion to applied regression. https:// cran.r-project.org/web/packages/car/ index.html. Accessed 30 Sep 2024. Fryxell, J. M., A. r. e. sinClAir, and G. CAuGhley. 2014. Wildlife ecology, con- servation, and management. John Wiley & Sons, Hoboken, NJ. GArner, d. l., h. B. underwood, and W. F. Porter. 1995. Use of modern infrared thermography for wildlife population surveys. Environmental Management 19: 233–238. GAsAwAy, w. C., s. d. duBois, d. J. reed, and S. J. hArBo. 1986. Estimating moose population parameters from aerial sur- veys. Biological Papers of the University of Alaska 22: 1–99. GelMAn, A., and J. hill. 2007. Data analysis using regression and multilevel/hierar- chical models. Cambridge University Press. https://www.ecfr.gov/current/title-14/part-107 https://www.ecfr.gov/current/title-14/part-107 https://cran.r-project.org/web/packages/car/index.html https://cran.r-project.org/web/packages/car/index.html https://cran.r-project.org/web/packages/car/index.html MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 162 Gentle, M., n. FinCh, J. sPeed, and A. PoPle. 2018. A comparison of unmanned aerial vehicles (drones) and manned helicop- ters for monitoring macropod popula- tions. Wildlife Research 45: 586–594. GreenBerG, S., T. Godin, and J. whittinGton. 2019. Design patterns for wildlife-related camera trap image analysis. Ecology and Evolution 9: 13706–13730. GuillerA-ArroitA, G., J. J. lAhoz-MonFort, d. i. MACkenzie, B. A. wintle, and M. A. MCCArthy. 2014. Ignoring imperfect detection in biological surveys is dan- gerous: a response to ‘fitting and inter- preting occupancy models’. PLOS One 9: e99571. hArris, r. B., M. AtAMiAn, h. FerGuson, and I. keren. 2015. Estimating moose abundance and trends in northeastern Washington state: Index counts, sightabilty models, and reducing uncer- tainty. Alces 51: 57–69. hAvens, K. J., and E. J. shArP. 2015. Thermal imaging techniques to survey and moni- tor animals in the wild: a methodology. Elsevier Science & Technology, Chantilly, USA. heit, d. r., w. ortiz-CAlo, M. k. P. Poisson, A. r. Butler, and R. J. Moll. 2024. Generalized nonlinearity in animal ecol- ogy: Research, review, and recommenda- tions. Ecology and Evolution 14: e11387. hinton, J. w., r. e. wheAt, P. sChuette, J. e. hurst, d. w. krAMer, J. h. stiCkles, and J. L. FrAir. 2022. Challenges and opportunities for estimating abundance of a low-density moose population. The Journal of Wildlife Management 86: e22213. hodGson, J. C., s. M. BAylis, r. Mott, A. herrod, and R. H. ClArke. 2016. Precision wildlife monitoring using unmanned aerial vehicles. Scientific Reports 6: 22574. hollAnd, J. d., d. G. Bert, and L. FAhriG. 2004. Determining the spatial scale of species’ response to habitat. BioScience 54: 227–233. iGlAy, r. B., l. r. Jones, J. A. elMore, k. o. evAns, s. sAMiAPPAn, M. B. PFeiFFer, and B. F. BlACkwell. 2024. Wildlife monitor- ing with drones: A survey of end users. Wildlife Society Bulletin 48: e1533. Jones, h., P. Pekins, l. kAntAr, i. sidor, d. ellinGwood, A. liChtenwAlner, and M. o’neAl. 2019. Mortality assessment of calf moose (Alces alces) during succes- sive years of winter tick (Dermacentor albipictus) epizootics in New Hampshire and Maine. Canadian Journal of Zoology 97: 22–30. kAntAr, L. E., and R. E. CuMBerlAnd. 2013. Using a double-count aerial survey to estimate moose abundance in Maine. Alces 49: 29–37. kéry, M., B. GArdner, and C. MonnerAt. 2010. Predicting species distributions from checklist data using site-occupancy models. Journal of Biogeography 37: 1851–1862. kéry, M., and J. A. royle. 2015. Applied hierarchical modeling in ecology: Analysis of distribution, abundance and species richness in R and BUGS / Volume 1, Prelude and static models. Elsevier, San Diego, California. kissell, R. E., Jr., and S. K. niMMo. 2011. A technique to estimate white-tailed deer Odocoileus virginianus density using vertical-looking infrared imagery. Wildlife Biology 17: 85–92. kretser, h., M. Glennon, A. whitelAw, A. hurt, k. PilGriM, and M. sChwArtz. 2016. Scat-detection dogs survey low density moose in New York. Alces 52: 55–66. krishnAn, B. S., L. R. Jones, J. A. elMore, S. sAMiAPPAn, k. o. evAns, M. B. PFeiFFer, B. F. BlACkwell, and R. B. iGlAy. 2023. Fusion of visible and ther- mal images improves automated detec- tion and classification of animals for drone surveys. Scientific Reports 13: 10385. lAMBA, A., P. CAssey, r. r. seGArAn, and L. P. koh. 2019. Deep learning for ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 163 environmental conservation. Current Biology 29: R977–R982. leCun, y., y. BenGio, and G. hinton. 2015. Deep learning. Nature 521: 436–444. levin, S. A. 1992. The problem of pattern and scale in ecology: The Robert H. MacArthur award lecture. Ecology 73: 1943–1967. linChAnt, J., J. lisein, J. seMeki, P. leJeune, and C. verMeulen. 2015. Are unmanned aircraft systems (UASs) the future of wildlife monitoring? A review of accom- plishments and challenges. Mammal Review 45: 239–252. link, W. A., and M. J. eAton. 2012. On thin- ning of chains in MCMC. Methods in Ecology and Evolution 3: 112–115. lonGMore, s. n., r. P. Collins, s. PFeiFer, s. e. Fox, M. Mulero-PázMány, F. BezoMBes, A. Goodwin, M. de JuAn ovelAr, J. h. knAPen, and S. A. wiCh. 2017. Adapting astronomical source detection software to help detect ani- mals in thermal images obtained by unmanned aerial systems. International Journal of Remote Sensing 38: 2623–2638. louhAiChi, M., M. M. BorMAn, and D. E. Johnson. 2001. Spatially located plat- form and aerial photography for docu- mentation of grazing impacts on wheat. Geocarto International 16: 65–70. MACFArlAne, C., and G. N. oGden. 2012. Automated estimation of foliage cover in forest understorey from digital nadir images. Methods in Ecology and Evolution 3: 405–415. MACkenzie, d. i., J. d. niChols, G. B. lAChMAn, s. droeGe, J. A. royle, and C. A. lAnGtiMM. 2002. Estimating site occupancy rates when detection proba- bilities are less than one. Ecology 83: 2248–2255. MAyer, M., t. voGler, G. Ausilio, t. kirChner, k. M. MAthisen, A. l. evAns, A. eriksen, M. heuriCh, v. nAuMov, P. wABAkken, and B. ziMMerMAnn. 2024. Drone-borne monitoring of moose. Research Report, Høgskolen i Innlandet. h t t p s : / / b r a g e . i n n . n o / i n n - x m l u i / handle/11250/3115853 MCMAhon, C. r., h. howe, J. vAn den hoFF, r. AlderMAn, h. BrolsMA, and M. A. hindell. 2014. Satellites, the all-seeing eyes in the sky: counting elephant seals from space. PLOS One 9: e92613. MCMAhon, M. C., M. A. ditMer, J. d. Forester. 2021a. Comparing unmanned aerial systems with conventional meth- odology for surveying a wild white- tailed deer population. Wildlife Research 49: 54–65. MCMAhon, M. C., M. A. ditMer, e. J. isAAC, s. A. Moore, and J. D. Forester. 2021b. Evaluating unmanned aerial systems for the detection and monitoring of moose in northeastern Minnesota. Wildlife Society Bulletin 45: 312–324. MiChAud, J.-s., n. C. CooPs, M. e. Andrew, M. A. wulder, G. s. Brown, and G. J. M. riCkBeil. 2014. Estimating moose (Alces alces) occurrence and abundance from remotely derived environmental indicators. Remote Sensing of Environment 152: 190–201. Millette, t. l., d. slAyMAker, e. MArCAno, C. AlexAnder, and L. riChArdson. 2011. AIMS-thermal - a thermal and high reso- lution color camera system integrated with GIS for aerial moose and deer census in northeastern Vermont. Alces 47: 27–37. Moll, r. J., M. Poisson, d. r. heit, h. Jones, P. J. Pekins, and L. kAntAr. 2022. A review of methods to estimate and monitor moose density and abundance. Alces 58: 31–49. MusAnte, A. r., P. J. Pekins, and D. L. sCArPitti. 2010. Characteristics and dynamics of a regional moose Alces alces population in the northeastern United States. Wildlife Biology 16: 185–204. NHFG: New Hampshire Fish and Game Department. 2015. 2015 state wildlife action plan. Annual Report, Concord, NH. https://brage.inn.no/inn-xmlui/handle/11250/3115853 https://brage.inn.no/inn-xmlui/handle/11250/3115853 MOOSE DETECTION VIA UASs ALCES VOL. 60, 2024 164 NHFG: New Hampshire Fish and Game Department. 2023. 2022 New Hampshire wildlife harvest summary. Annual Report, Concord, NH. NOAA: National Oceanic and Atmospheric Administration. 1991. NOWData: Gorham Airport, NH, USA. Daily Climate Normals, National Weather Service. . Accessed 9 Dec 2024. northruP, J. M., and B. D. GerBer. 2018. A comment on priors for Bayesian occu- pancy models. PLOS One 13: e0192819. ooMs, J. 2024. magick: advanced graphics and image-processing in R. . Accessed 30 Sep 2024. oyster, J. h., i. n. keren, s. J. k. hAnsen, and R. B. hArris. 2018. Hierarchical mark-recapture distance sampling to esti- mate moose abundance. The Journal of Wildlife Management 82: 1668–1679. Peters, w., M. heBBlewhite, k. G. sMith, s. M. weBB, n. weBB, M. russell, C. stAMBAuGh, and R. B. Anderson. 2014. Contrasting aerial moose population estimation methods and evaluating sightability in west-central Alberta, Canada. Wildlife Society Bulletin 38: 639–649. Poole, k. G., G. MowAt, and D. PritChArd. 1999. Using GPS and GIS for naviga- tion and mark-recapture for sightability correction in moose inventories. Alces 35: 1–10. QuAyle, J. F., A. G. MAChutChon, and D. N. Jury. 2001. Modeling moose sightabil- ity in south-central British Columbia. Alces 37: 43–54. r Core teAM. 2024. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. rAhMAn, D. A., and A. A. A. F. rAhMAn. 2021. Performance of unmanned aerial vehicle with thermal imaging, camera trap, and transect survey for monitoring of wildlife. IOP Conference Series: Earth and Environmental Science 771: 012011. reMPel, r., P. elkie, A. rodGers, and M. GluCk. 1997. Timber-management and natural-disturbance effects on moose habitat: landscape evaluation. The Journal of Wildlife Management 61: 517–524. ruPreCht, J. s., d. n. koons, k. r. hersey, n. t. hoBBs, and D. R. MACnulty. 2020. The effect of climate on population growth in a cold-adapted ungulate at its equatorial range limit. Ecosphere 11: e03058. sAlA, o. e., F. stuArt ChAPin, iii, J. J. ArMesto, e. Berlow, J. BlooMField, r. dirzo, e. huBer-sAnwAld, l. F. huenneke, r. B. JACkson, A. kinziG, r. leeMAns, d. M. lodGe, h. A. Mooney, M. oesterheld, n. l. PoFF, M. t. sykes, B. h. wAlker, M. wAlker, and D. H. wAll. 2000. Global Biodiversity Scenarios for the Year 2100. Science 287: 1770–1774. sChultz, e. A., n. ellison-neAry, M. r. BoudreAu, G. M. street, l. r. Jones, k. o. evAns, and R. B. iGlAy. 2024. On the move: Influence of animal movements on count error during drone surveys. Ecology and Evolution 14: e70287. seyMour, A. C., J. dAle, M. hAMMill, P. n. hAlPin, and D. W. Johnston. 2017. Automated detection and enumeration of marine wildlife using unmanned air- craft systems (UAS) and thermal imag- ery. Scientific Reports 7: 45127. solBerG, e. J., B.-e. sAether, o. strAnd, and A. loison. 1999. Dynamics of a har- vested moose population in a variable environment. Journal of Animal Ecology 68: 186–204. sPAAn, d., C. Burke, o. MCAree, F. Aureli, C. e. rAnGel-riverA, A. hutsChenreiter, s. n. lonGMore, P. r. MCwhirter, and S. A. wiCh. 2019. Thermal infrared imaging from drones offers a major advance for spider monkey surveys. Drones 3: 34. https://www.weather.gov/wrh/climate https://www.weather.gov/wrh/climate https://cran.r-project.org/web/packages/magick/index.html https://cran.r-project.org/web/packages/magick/index.html https://cran.r-project.org/web/packages/magick/index.html ALCES VOL. 60, 2024 MOOSE DETECTION VIA UASs 165 steinhorst, R. K., and M. D. sAMuel. 1989. Sightability adjustment methods for aerial surveys of wildlife populations. Biometrics 45: 415–425. street, G. M., A. R. rodGers, and J. M. Fryxell. 2015. Mid-day temperature variation influences seasonal habitat selection by moose. The Journal of Wildlife Management 79: 505–512. su, Y.-S., and M. yAJiMA. 2024. R2JAGS: using R to run “JAGS.” . Accessed 30 Sep 2024. sullivAn, F. B., A. G. hunsAker, M. w. PAlACe, and J. M. JACoBs. 2023. Evaluating the effects of UAS flight speed on LiDAR snow depth estimation in a heterogeneous landscape. Remote Sensing 15: 5091. tAPe, k. d., d. d. Gustine, r. w. ruess, l. G. AdAMs, and J. A. ClArk. 2016. Range expansion of moose in arctic Alaska linked to warming and increased shrub habitat. PLOS ONE 11: e0152636. teitelBAuM, C. s., A. P. k. sirén, e. CoFFel, J. r. Foster, J. l. FrAir, J. w. hinton, r. M. horton, d. w. krAMer, C. lesk, C. rAyMond, d. w. wAttles, k. A. zeller, and T. L. Morelli. 2021. Habitat use as indicator of adaptive capacity to climate change. Diversity and Distributions 27: 655–667. tyre, A. J., B. tenhuMBerG, s. A. Field, d. nieJAlke, k. PArris, and H. P. PossinGhAM. 2003. Improving precision and reducing bias in biological surveys: estimating false-negative error rates. Ecological Applications 13: 1790–1801. wAnG, d., Q. shAo, and H. yue. 2019. Surveying wild animals from satellites, manned aircraft and unmanned aerial systems (UASs): a review. Remote Sensing 11: 1308. wiens, J. A. 1989. Spatial scaling in ecology. Functional Ecology 3: 385–397. zABel, F., M. A. FindlAy, and P. J. C. white. 2023. Assessment of the accuracy of counting large ungulate species (red deer Cervus elaphus) with UAV- mounted thermal infrared cameras during night flights. Wildlife Biology 2023: e01071. ZmarZ, a., m. rodZewicZ, m. dąbski, i. kArszniA, M. korCzAk-ABshire, and K. J. ChwedorzewskA. 2018. Application of UAV BVLOS remote sensing data for multi-faceted analysis of Antarctic eco- system. Remote Sensing of Environment 217: 375–388. zuur, A. F., e. n. ieno, and C. S. elPhiCk. 2010. A protocol for data exploration to avoid common statistical problems. Methods in Ecology and Evolution 1: 3–14. https://cran.r-project.org/web/packages/R2jags/index.html https://cran.r-project.org/web/packages/R2jags/index.html https://cran.r-project.org/web/packages/R2jags/index.html _Hlk208300554