Very interesting and potentially important analysis of the best way of carrying out aerial animal counts. Could have profound implications for our understanding of numbers for large mammal populations – especially elephants. Time to try this in Botswana and the KAZA (Kavango–Zambezi Transfrontier Conservation Area) region. KS

Zoological Society of London 

Cameras replace human observers in multi‐species aerial counts in Murchison Falls, Uganda

First published:15 March 2020

Funding Information:

This work was funded by TOTAL E&P Uganda BV (TEPU), supported by Flora & Fauna International operating under a TEPU service contract to conduct aerial wildlife surveys of MFPA. Save‐the‐Elephants of Nairobi, Kenya, generously sponsored further analysis of results for production of this paper. Vulcan Inc. of Seattle, USA, kindly authorized the use of the 2014 Great Elephant Census (GEC) survey dataset for Murchison for our comparative analysis

Abstract

Wildlife counts in Africa and elsewhere are often implemented using light aircraft with ‘rear‐seat‐observer’ (RSO) counting crews. Previous research has indicated that RSOs often fail to detect animals, and that population estimates are therefore biased. We conducted aerial wildlife surveys in Murchison Falls Protected Area, Uganda, in which we replaced RSOs with high‐definition ‘oblique camera count’ (OCC) systems. The survey area comprises forests, woodlands and grasslands. Four counts were conducted in 2015–2016 using a systematic‐reconnaissance‐flight (SRF) strip‐transect design. Camera inclination angles, focal lengths, altitude and frame interval were calibrated to provide imaged strips of known sample size on the left and right sides of the aircraft. Using digital cameras, 24 000 high‐definition images were acquired for each count, which were visually interpreted by four airphoto interpreters. We used the standard Jolly II SRF analysis to derive population estimates. Our OCC estimates of the antelopes – hartebeest, Uganda kob, waterbuck and oribi – were, respectively, 25%, 103%, 97% and 2100% higher than in the most recent RSO count conducted in 2014. The OCC surveys doubled the 2014 RSO estimate of 58 000 Uganda kob to over 118 000. Population size estimates of elephants and giraffes did not differ significantly. Although all four OCC buffalo estimates were higher than the RSO estimates – in one count by 60% – these differences were not significant due to the clumped distribution and high variation in herd sizes, resulting in imprecise estimation by sampling. We conclude that RSO wildlife counts in Murchison have been effective in enumerating elephants and giraffe, but that many smaller species have not been well detected. We emphasize the importance of 60 years of RSO‐based surveys across Africa, but suggest that new imaging technologies are embraced to improve accuracy.

Introduction

Wildlife surveys over large and remote wilderness areas in Africa, America and Australia are often conducted using light aircraft with ‘rear‐seat‐observer’ (RSO) crews who count within defined ‘strip‐transects’ (Caughley 1977; Norton‐Griffiths 1978; Grimsdell and Westley 1981; Gasaway et al. 1986; PAEAS, 2014), or record distance of animals from the aircraft path in the ‘line‐transect’ method (Burnham et al. 1985; Pollock and Kendall 1987; Samuel et al. 1987; Hone 1988; Buckland et al. 2004). It has long been recognized that RSOs may fail to detect animals, resulting in negatively biased population estimates (Caughley 1974; Cook and Jacobson 1979; Pollock and Kendall 1987; Graham and Bell 1989; Jachmann 2002; Tracey et al. 2008; Jacques et al. 2014; Lee and Bond 2016). Animals might not be detected either because they are ‘unavailable for detection’, being for example hidden in dense vegetation cover or underwater (Bayliss and Yeomans 1989; Marsh and Sinclair 1989; Jachmann 2001; Mackie et al. 2013; Jacques et al. 2014) or because they available but ‘overlooked’ by the RSOs (Fleming and Tracey 2008). RSO detection of ‘available’ animals depends on a range of ‘environmental factors’ such as animal size, group size, vegetation cover, species coloration, reaction to the aircraft, occurrence in multi‐species assemblages; ‘survey factors’ such as flying height, counting strip‐width and sun angle; and ‘observer factors’ such as experience and level of fatigue (Caughley et al. 1976; Anderson and Lindzey 1996; Jachmann 2002; Melville et al. 2008; McConville et al. 2009; Wal et al. 2011; Ransom 2012; Griffin et al. 2013; Jacques et al. 2014; Strobel and Butler 2014; Lubow and Ransom 2016; Schlossberg et al. 2016; Schlossberg et al. 2017).

For the last 60 years in East Africa, the RSO‐based ‘systematic‐reconnaissance‐flight’ (SRF) strip‐transect technique, coupled with the ‘Jolly II analysis for unequal sized sample units’ (Jolly 1969), has been the standard procedure for counting wildlife and livestock at the local and the national level (Andere 1981; Ottichilo et al. 2000). In the traditional SRF technique, the aircraft follows a systematic flight pattern of transects, usually aligned to a spatial grid‐system, while RSOs count animals within sample strips defined on each side of the aircraft (Jolly 1969; Norton‐Griffiths 1978; Gasaway et al. 1986). In Kenya, a national programme of surveys using ‘SRF‐Jolly II’, launched in 1977, has provided a unique 40‐year record for determining trends in wildlife and livestock populations (Ogutu et al. 2016). In Tanzania where the SRF technique was developed in the late 1960s for mapping seasonal distributions of wildlife in Serengeti (Maddock 1979; Norton‐Griffiths 1981), SRF surveys inform policymakers of the status of wildlife populations, and particularly of elephants (TAWIRI, 2010b). At the continental scale, SRF results are compiled to determine the overall population status of the African elephant (Thouless et al. 2016).

The SRF technique is therefore ubiquitous, but concern is often raised that SRF‐derived population estimates are not precise (have high margins of error), or fluctuate wildly, or differ significantly (usually being lower) with estimates derived from ground counts which are assumed to more accurately reflect the population (Grimsdell and Westley 1981; de Leeuw et al. 1998; Jachmann 2002; Ferreira and Van Arde 2009; Lee and Bond 2016; Greene et al. 2017; Reilly et al. 2017). In the context of SRFs, many studies have been conducted to improve RSO counting, from determining optimum flying heights and strip widths (Pennycuick and Western 1969; Caughley 1974), to RSO use of cameras for photographing large herds for later counting (Sinclair 1973; Norton‐Griffiths 1974), to double‐observation techniques with two RSOs on each side of the aircraft (Magnusson et al. 1978; Caughley and Grice 1982; Pollock et al. 2006; Griffin et al. 2013; Schlossberg et al. 2016). In recent initiatives to standardize procedures, guidelines have been developed that prescribe such fundamental aspects as eyesight and counting tests for RSOs and the maximum counting periods ‘on‐transect’ (Frederick et al. 2010; Craig 2012; PAEAS, 2014). Studies to determine a ‘probability of detection’ for SRFs have largely focussed on comparisons with ground counts (Stelfox and Peden 1981; Jachmann 2002; Greene et al. 2017), but results are usually species and survey specific. Elsewhere, for example, in the US, such biases are also ascertained when marked or radio‐tagged animals known to be within the viewing field of the RSOs were not detected (Rice et al. 2009; Wal et al. 2011; Jacques et al. 2014; Lubow and Ransom 2016), essentially the approach that ‘we‐know‐they‐are‐there, but‐you‐did‐not‐see‐them’. This requires the availability of many tagged animals for a meaningful sample, and there is no evidence that this approach has been tried in the context of an SRF survey in Africa.

SRFs over large areas are conducted using fixed‐wing aircraft, since helicopters do not have the endurance for large surveys, and are usually unavailable or unaffordable. Aircraft must operate at speed well above the stall to remain safe, and hence ground speeds of 160–180 km.hr−1 are usually prescribed (Craig 2012; PAEAS, 2014). At this speed, any particular component of the sample‐strip scene – grassland, woodland, open glade, patch of wetland – remains in the RSO’s view for no more than 5 seconds. In multi‐species SRFs, within this short time, the RSO must detect the species (for example, wildebeest, zebra, gazelle); prioritize which species must be counted first; count the animals; possibly photograph the herd; ‘subtize’ (short‐term memorize) the estimate and image number (Fleming and Tracey 2008); record both of these on voice recorder or call estimates to front‐seat‐observer (FSO, the ‘recorder’) (Craig 2012; PAEAS, 2014); repeat this process species for species 2 and 3 etc. Calling to the FSO might coincide with a call‐out of the opposite RSO in the aircraft who has also seen animals, causing confusion and distraction to all parties. Observing is characterized by long periods of nothing (where the mind may wander), punctuated with bouts of frenetic counting where the observer may be overwhelmed with large herds or congregations of many species. Often multi‐species counts are focussed on 1–3 priority species, for example, elephant, buffalo and giraffe; ‘supplementary species’, for example, the antelopes are added because program managers argue that “it costs us US $ 700 an hour to keep this plane in the air and we want to collect as much data as possible”. Tired RSOs do not prioritize supplementary species, with the result that these count data are of low quality.

For many years researchers have suggested that to reduce bias, cameras could replace observers in large‐area counts (Leedy 1948; Siniff and Skoog 1964; Caughley 1974). However, in early attempts using analogue camera systems for large‐area surveys, the logistics of handling, geo‐referencing, processing and interpreting huge volumes of analogue imagery inevitably proved challenging (Terletzky 2013). Early examples include oblique and ‘aerial‐point‐sampling’ (APS) surveys of wildebeest and buffalo in Serengeti, Tanzania, (Norton‐Griffiths 1973; Sinclair 1973), and caribou in Canada (Couturier et al. 1994). The use of digital cameras has greatly enhanced the scope for large area surveys, but development has been relatively slow since, inevitably, this still requires the use of expensive aircraft and the visual interpretation of thousands of images. However, where the target wildlife population runs into millions, or survey areas are very remote, digital camera APS surveys are the only possible way to count; they are now periodically conducted in Serengeti (TAWIRI, 2010a; Hopcraft et al. 2015), with a recent further application in Mongolia (Norton‐Griffiths et al. 2015).

High‐resolution digital cameras are now cheap, available and small, and data storage media have a capacity for many thousands of images. It is now possible to test RSO performance with parallel digital camera systems that are inclined at the same angle as RSO‐viewing, and image the same strip. Recent cross‐comparisons in simultaneous RSO and OCC counting have been conducted for narwhal in Greenland (Monodon monoceros) (Bröker et al. 2019) and kangaroos in Australia (Lethbridge et al. 2019) where thermal image‐based estimates of kangaroo density were 30% higher than RSO estimates. In Kenya, an RSO‐based SRF was run concurrently with high‐resolution camera systems in a multi‐species count over a large protected area, and it was found that RSOs missed, for example, 60% of giraffe and 66% of the large antelopes (Lamprey et al. 2019).

In this paper, we report on the earliest known experiment in Africa in which an ‘oblique‐camera count’ (OCC) system entirely replaced RSOs in a systematic reconnaissance flight (Lamprey 2016); the study established the later methods for the Kenya surveys indicated above. In Uganda’s Murchison Falls Protected Area (MFPA), we tested the hypothesis that high‐resolution camera systems, set up obliquely to replicate RSO strip‐sample counting, would generate higher and more consistent population estimates than those derived from RSOs. To achieve this, we acquired continuous imagery along SRF transects, and interpreted this imagery for species and numbers in the laboratory. We then compared our estimates with those derived from recent RSO‐based counts of MFPA.

In MFPA, Uganda kob (Kobus kob ssp. thomasi) provide a special case for investigating counting performance. The Uganda kob (pl. kob) is the national emblem species, being the main feature of the country’s coat of arms and also the logo of the Uganda Wildlife Authority. Uganda kob occur in highly clumped aggregations around territorial breeding grounds (or ‘leks)’ (Balmford 1992; Deutsch 1994) that persist for years. Consequently, aerial sample counts where transects are not aligned in similar orientation give wildly varying estimates for this species (Modha and Eltringham 1976). This species was heavily impacted by poaching in the 1980s, with RSO‐based population estimates of approximately 5300 in 1995 (Lamprey and Michelmore 1995; Sommerlatte and Williamson 1995) and 7458 in 1999 (Lamprey 2000). Since that time, with improved management of MFPA, this population has increased exponentially, with 9315 estimated in 2005 (Rwetsiba and Wanyama 2005), 36 234 in 2012 (Rwetsiba et al. 2012), and 58 313 in 2014 (Wanyama et al. 2014). With kob numbers evidently increasing so rapidly, the ability of RSOs to effectively count them is called into question…

Discussion

Our research shows that oblique camera‐counts generate population estimates that are significantly higher and more consistent than those derived from RSO counts. The potential impact of camera‐counts on nation‐wide wildlife inventories is high. Our surveys increase the national population estimate for Uganda kob by 77%, from 77 759 (UWA, 2015) to 137 736. With improved protection, high rainfall and low levels of predation, kob in Block 1 have increased to a density of 78 kob.km−2, well beyond the ‘record’ densities of 45 kob.km−2 in Toro Game Reserve in the 1960s (Beuchner 1974), and probably the highest ever recorded in Uganda (Modha and Eltringham 1976). The MFPA Uganda kob increase is probably unprecedented for any wild antelope population in recent times, with similarities to the exponential increase of wildebeest in the Serengeti in the 1970s and 80s (Sinclair 1979; Hopcraft et al. 2015) and the George River caribou herd in Canada from 1960–1990 (Messier et al. 1988) ‐ the latter sadly in catastrophic decline to just 8000 of the estimated population of 800,000 in the 90s (Canada GNL 2018; Romea, 2018).

With regard to other species in MFPA, the Lelwel hartebeest, listed as Endangered by IUCN (IUCN Red List, 2017) and which was reduced in South Sudan from over 50 000 in the 1960s to just 1100 in 2007 (Fay et al. 2007), is shown in the camera‐counts as having a relatively healthy and increasing population in MFPA; this is probably the most important population of this sub‐species globally. Elsewhere, for species that aggregate in large numbers, camera‐counts, might also positively alter national species inventories, for example, wildebeest in Maasai Mara in Kenya (Bhola et al. 2012), white‐eared kob in South Sudan (Fay et al. 2007) and Saiga antelope in Kazakhstan (McConville et al. 2009). In the Serengeti, with over one million wildebeest (Hopcraft et al. 2015), periodic censuses of wildebeest and other migratory wildlife are now entirely conducted by vertical aerial‐point‐sampling (Norton‐Griffiths 1973; TAWIRI‐CIMU, n.d.).

In addition to more precise counting, camera‐counts have a number of key advantages over RSO‐based surveys. Firstly, the imagery provides the full sample, frozen in time for verification and reanalysis of species numbers and strip‐widths, and for further exploration of factors that determine visibility and distributions (Jacques et al. 2014; Schlossberg et al. 2016; Ndaimani et al. 2017). Secondly, while image interpreters need enthusiasm and patience, they do not need to be experienced to deliver consistent results. During the study, two interpreters departed and were replaced with new graduates who, after short training and mentoring, performed equally well. These findings are reflected in other image interpretation studies (Erwin 1982; Frederick et al. 2003) where untrained interpreters outperformed advanced remote sensing techniques (Terletzky and Ramsey 2016). Thirdly, there are safety dividends in excluding RSO passengers in low‐level surveys and transiting to image‐based methods; SRFs are operated at low‐levels, where bird strikes, power lines and violent turbulence (especially in mountainous area) are hazards, and where engine failure is potentially catastrophic. Without RSOs who need to count at altitudes of 350 ft or lower (PAEAS, 2014), aircraft with camera‐count systems can fly higher and still deliver high‐resolution imagery for counting; currently, the optimum OCC height above ground level (HAGL) is set at 600 ft (Lamprey 2018).

The disadvantage of the OCC approach in its current stage is the high volume of imagery generated, and associated labour costs for interpretation. In assessing costs, a wildlife SRF budget is divided into three components; (1) the technical components of design, oversight, data management, data checking, analysis, mapping (GIS work) and reporting; (2) the cost of operating the aircraft; (3) the cost of the data acquisition, whether this is by RSOs (for flight allowances, salaries, accommodation in the field) or airphoto interpreters (for remuneration costs). OCC surveys do not increase components 1 or 2, although data checking might add a small increment. This might be offset by a cheaper aircraft, which could for example be a 2‐seat aircraft or even microlight, since we no longer need to carry observers in a 4‐ or 6‐seat aircraft. The larger cost of the OCC method is in component 3, where current calculations indicate that the OCC method costs approximately 25% more than RSO‐based data acquisition. Given the offset with aircraft costs, the budget for a OCC survey, carried out at the same sampling intensity, is similar to that of an equivalent RSO survey. For the 12% sample for MFPA described in our study, the costs per sampled km2 (total 605 km2) are indicated as US $64.km−2, with 30% of this cost for image interpretation. The generation of estimates and distribution maps for each MFPA survey was achieved in three months, which was acceptable for the wildlife managers of the area.

Costs may also come down with the use of UAVs for imaging surveys. While great progress has been made in their use for wildlife surveys, endurance limitations restrict their use to small areas (Wang et al. 2019). In West Africa, for example, a UAV was successfully used for an elephant SRF of a small reserve of 940 km2, but limitations on range and reliability meant that the exercise took many days to complete (Vermeulen et al. 2013). It was concluded that the exercise cost 10 times as much as with using a light aircraft. Today, a UAV may not complete more than one 60 km transect in MFPA, but in time, with improvements in reliability and endurance, UAVs will be routinely used for large area counts.

A major constraint to OCC efficiency is that, in emulating an RSO count, there are long stretches where no animals are encountered. Thus, for example, in the MU3 survey, 2165 out of 23 927 images (9.3%) captured target animals, and of these just 25, or 0.1% of the total, contained elephants. An important immediate avenue for investigation is to sub‐sample the dataset to test if precision can be maintained at lower volumes of imagery (Norton‐Griffiths et al. 2015), and to determine how to filter out true‐negatives, the images with nothing in them. New techniques in machine learning now offer the possibility of species identification and enumeration, (Sirmacek et al. 2012; Rey et al. 2017; Eikelboom et al. 2019; Tabak et al. 2019), but it will take some time for experimental techniques to be operationalized for full‐scale SRFs. Although research has focussed on specific species and sites, an immediate need is to derive AI systems which can simply filter out images with ‘something‐of‐interest’ from the > 80% of the true negatives. This leaves human interpreters with a greatly reduced workload.

Improvements in visible spectrum and thermal IR aerial sensors, both in manned aircraft and UAVs, now offer new opportunities for animal counts (Terletzky 2013; Lethbridge et al. 2019), with the highest successes in detecting objects with high contrast, for example, seals or penguins on ice‐flows (Conn et al. 2014; McMahon et al. 2014; Borowicz et al. 2018). Researchers are also now turning to high‐resolution satellite sensors such as Ikonos, GeoEye‐1, and WorldView‐3 (Laliberte and Ripple 2003; Fretwell et al. 2017; Xue et al. 2017), especially for counts of waterfowl and seabirds. While there have been some advances in counting animals in the open savannas in Africa (Yang et al. 2014), it will take some years before these techniques can be applied in complex wooded savannah environments.

The development of remote‐sensing methods for wildlife counts are moving forward rapidly, but at the same time the traditional RSO‐based SRF count remains valuable and relevant to providing long‐term wildlife trends to policymakers (Ogutu et al. 2016). To determine the status of such key species as elephants, it is essential that efforts are continued to improve and standardize RSO‐based counts (Jachmann 2002; Craig 2012; PAEAS, 2014). Where information is needed on the distribution and abundance of multiple species in a landscape, the OCC method described in this paper is a significant step in the evolution of more accurate and automated wildlife counting.