How do you find the Green Sheep? A critical review of the use of remotely sensed imagery to detect and count animals

Tracey Hollings*1, Mark Burgman1,2, Mary van Andel3, Marius Gilbert4,5, Timothy Robinson6, Andrew Robinson1

1 Centre of Excellence for Biosecurity Risk Analysis, School of Biosciences, University of Melbourne, Melbourne Australia 3010

2 Centre for Environmental Policy, Imperial College London, U.K.

3 Ministry for Primary Industries, Wellington, New Zealand 6140

4 Spatial Epidemiology Lab., Université Libre de Bruxelles, Brussels, Belgium

5Fonds National de la Recherche Scientifique, Brussels, Belgium

6 Food and Agricultural Organisation of the United Nations, Rome, Italy

*Corresponding author: Phone: +61 3 8344 0071; Email: tracey.hollings@unimelb.edu.au

Word count: 6,982

Abstract

1. Animal abundance data are essential for endangered species conservation, tracking invasive species spread, biosecurity, agricultural applications and wildlife monitoring; however obtaining abundance data is a perennial challenge. Recent improvements in the resolution of remotely sensed imagery, and image processing tools and software have facilitated improvement of methods for the detection of individual, generally large-bodied animals. The potential to monitor and survey populations from remotely sensed imagery is an exciting new development in animal ecology.

2. We review the methods used to analyse remotely sensed imagery for their potential to estimate the abundance of wild and domestic animal populations by directly detecting, identifying and counting individuals.

3. Despite many illustrative studies using a variety of methods for detecting animals from remotely sensed imagery, it remains problematic in many situations. Studies that demonstrated reasonably high accuracy using automated and semi-automated techniques have been undertaken on small spatial scales relative to the geographic range of the species of interest and/or in homogenous environments such as sea ice. The major limitations are the relatively low accuracy of automated detection techniques across large spatial extents, false detections, and the cost of high resolution data.

4. Future developments in the analysis of remotely sensed data for population surveys will improve detection capabilities, including the advancement of algorithms, the crossover of software and technology from other disciplines, and improved availability, accessibility, cost and resolution of data.

Key words

Satellite imagery, remote sensing, population survey, animals, wildlife, agriculture, detection

Introduction

Animal abundance information is essential for many applications including endangered species conservation, tracking invasive species spread, agriculture, and wildlife monitoring. Yet, traditional, ground-based population surveys can be cost-prohibitive, biased, logistically challenging, and time-consuming, particularly in large and hard to access areas (LaRue et al. 2011; Lynch et al. 2012; Fretwell, Staniland & Forcada 2014; LaRue et al. 2014; Oishi & Matsunaga 2014; Terletzky & Ramsey 2016). Remote sensing (RS) technologies may offer a viable alternative to traditional population survey techniques. Although animal population estimation from remotely sensed imagery were proposed more than 40 years ago (e.g. Kadlec & Drury 1968; Leonard & Fish 1974; Löffler & Margules 1980), studies generally have used indirect measures or surrogates based on an animal’s ecology to predict population abundances and distributions, or environmental proxies related to their presence (e.g. Lynch et al. 2012; LaRue et al. 2014; Robinson et al. 2014). For example, Löffler and Margules (1980) were able to approximate wombat distributions (Lasiorhinus latifrons) by detecting burrows, Velasco (2009) detected active marmot (Marmota siberica) mounds, Williams and Dowdeswell (1998) used the unique spectral signatures of vegetation attributed to altered soils below seabird nesting colonies to indicate the presence of seabirds, and studies have estimated penguin populations from guano-stained areas (e.g. Fretwell et al. 2012; Lynch et al. 2012; LaRue et al. 2014).

Recently, the development and proliferation of new technologies, and the greater resolution and access to RS data have improved the direct detection of individual organisms (e.g. Laliberte & Ripple 2003; Groom et al. 2011; LaRue et al. 2011; Yang 2012; Fretwell, Staniland & Forcada 2014; McMahon et al. 2014; Stapleton et al. 2014). This has enhanced the potential to estimate animal abundance by counting individuals (Pettorelli et al. 2014). Estimating populations using direct animal detection techniques from RS imagery is becoming increasingly common, and proof-of-concept studies using automated and semi-automated techniques have been effective with moderate and high resolution satellite imagery (e.g. McNeill et al. 2011; Groom et al. 2013; Conn et al. 2014). Estimating populations by detecting individual animals has been used mainly in the context of wildlife conservation, in particular for inaccessible or remote areas (e.g. Barber-Meyer, Kooyman & Ponganis 2007, Emperor penguins, Aptenodytes forsteri), and recently for domestic species including cattle and horses (Terletzky & Ramsey 2014).

Remotely sensed imagery has several benefits over ground-based animal surveys including large spatial coverage (e.g. global), regular updating and reassessment, and faster data procurement. RS technologies provide an opportunity to obtain data where previously it may not have existed, and data is imperative for ecological and conservation programs (Pettorelli et al. 2014). It also provides a permanent record of study sites at an instant in time, is unobtrusive, and requires less human intervention (Pettorelli et al. 2014; Yang et al. 2014; Terletzky & Ramsey 2016). Obtaining repeat imagery for replicating population surveys is also relatively straightforward with revisit intervals of some satellites only days apart. With continued developments in image processing, remotely sensed data can augment or potentially even reduce the need for data gathered by traditional, ground-based population surveys.

In this paper, we examine methods for analysing RS imagery to estimate the abundance of wild and domestic animals by directly detecting, identifying and counting individuals. We review recent studies for methods and data requirements, and the practical and operational considerations of using RS technologies in estimating animal abundance. We consider the advantages of better resolution and lower cost imagery, greater computing power, and more advanced statistical algorithms and programming.

Remote Sensing Data

Remote sensing captures electromagnetic (EM) waves from an object or scene remotely, usually via an aircraft or satellite. The EM spectrum is classified based on wavelengths into optical (0.4-14µm) or microwave (`1mm-1m), and different technologies are required to detect them (Turner et al. 2003). Research objectives drive spatial and spectral resolution requirements, which affect data costs. Spectral resolution refers to the width of the EM bands that are recorded; panchromatic imagery records in black and white, multispectral imagery captures many bands, including the visible light range of red, green and blue (RGB), and hyperspectral data collects hundreds of bands. Multispectral and hyperspectral data also include near-infrared, thermal and infrared wavelengths and provide more spectral information for discriminating animals and background (Turner et al. 2003; Chrétien, Théau & Ménard 2016). Spatial resolution refers to the amount of area each pixel represents within an image. Low spatial resolution has fewer pixels, and a pixel will represent a bigger area relative to higher spatial resolutions. This can affect the ability to distinguish two objects close together, or affect detectability if the object of interest is sufficiently smaller than the pixel size.

Three common methods for the acquisition of RS imagery are: satellites, unpiloted aerial vehicles (UAV), and aerial imagery from light aircraft. The availability, constraints, limitations and costs for each data collection method differ substantially. Satellite data offers global coverage, a range of spatial and spectral resolutions, and is available from commercial satellites. Worldview4 (launched in 2016) currently has the highest resolution imagery available: panchromatic imagery at 30cm resolution, and multispectral data at 1.24m. Cost to purchase 50cm spatial resolution satellite imagery ranges from USD$14-29 per km2 depending on spectral resolution and data age (Landinfo Worldwide Mapping LLC 2017). Minimum ordering requirements can considerably increase costs, but will likely reduce over time. Satellite data is ideal for large and remote areas, for example in Antarctica (Barber-Meyer, Kooyman & Ponganis 2007), and/or where spatial resolution doesn’t need to be greater than 50cm.

Aerial imagery from aircraft and UAV has several advantages over satellite derived images; flights can be timed for certain events, including specifying time between repeat area coverage, or to avoid cloud cover, and vary altitudes, and they support faster uptake of technological and sensor developments. Perhaps the greatest advantage of aerial imagery is that images can be collected with significantly higher spatial resolutions, up to 2.5cm (e.g. Aerometrex 2016). However aerial imagery is restricted to areas within reach of available aircraft, for example, access to runways or fuel limits for aeroplanes, or operating distance from users for UAV which may range to tens of kilometres. This can limit remote access, or increase deployment costs. Aerial imagery costs will depend on several factors including the location and size of the area to be mapped, availability of operators, and data resolution. UAV’s encompass a broad range of sizes from quad-copters to full size planes and therefore vary in cost and necessary skills. They are a relatively new technology for obtaining aerial imagery and are beginning to be used for animal detection (e.g. Chrétien, Théau & Ménard 2016). A UAV’s ability to capture high resolution imagery is limited by the sensors it can carry: a small payload means small, often lower quality sensors, and requires more passes to cover the same area as a fixed wing aircraft (Deuter 2016). UAVs tend to have a short line of sight from operator and length of a pass, can be expensive for labour and transportation, and are affected by weather, particularly wind (Chrétien, Théau & Ménard 2016; Deuter 2016). Continuing technological developments in UAV’s will overcome many of these issues and transform the use of UAVs in spatial ecology (Anderson & Gaston 2013).

Methods for the direct detection of individual animals from remotely sensed imagery

Many methods have been explored for directly detecting and counting individual, generally large-bodied (>10kg), animals from RS imagery (Table 1). Manual methods have been the most thoroughly investigated over the past four decades (e.g. Kadlec & Drury 1968; Leonard & Fish 1974; Löffler & Margules 1980). Automated and semi-automated techniques have developed in the past 15 years, arising concomitantly with technological and computer advancements. Different methods show varying results based on the type of imagery, size of the study area, and the number of animals, highlighting that application and context also affect results. For example, increasing numbers of animals in a survey area will increase the chances of detection (LaRue, Stapleton & Anderson 2017), but animals congregating close together has been reported to decrease the accuracy of population counts (Terletzky 2013). Most studies report results as correlations with ground surveys or as the proportion of animals detected from a known population. Errors of omission and commission arise where known animals are not detected or objects are incorrectly identified as animals, respectively. Incorrect population estimates of animals can have serious implications for the application of the data, e.g. quotas for harvesting could be set too high based on overestimating population size, or there may be implications for the conservation of a species, particularly if population trends are not accurately identified.

Manual methods

Manual counts use analysts to detect and identify animals in RS imagery and have been used for decades (e.g. Kadlec & Drury 1968; Leonard & Fish 1974; Löffler & Margules 1980), but can be time consuming, subjective, and costly due to the labour costs for image preparation/analysis (Stapleton et al. 2014; Terletzky & Ramsey 2014). Manual counts are extremely labour and time-intensive to apply successfully over spatial extents larger than a few square kilometres. In restricted areas, manual counts from RS imagery can be significantly correlated with more traditional survey techniques, including ground counts (r=0.98, n=1000, LaRue et al. 2011; r-squared=0.91, n=<2000; McMahon et al. 2014). RS imagery analysis comparisons with ground-truthed studies often report underestimation of population sizes because some animals are hidden from observers (Jachmann 2002; McMahon et al. 2014). Environmental homogenity, body size and the ability to differentiate individuals from their background all affect accuracy (Jachmann 2002; Terletzky & Ramsey 2014). Observer confidence and experience can also affect detection and misclassification rates (Stapleton et al. 2014). In general, manual counts in isolation are error-prone and expensive, especially over large areas, which is true of many traditional survey techniques, creating an imperative for automated and semi-automated techniques for surveying large geographic areas.

Crowd sourcing citizen science

One alternative approach for manually counting animals from remotely sensed imagery is to use crowdsourcing platforms. Crowdsourcing as outlined by Papadopoulou and Giaoutzi (2014) is “nowadays extensively used to describe a process, including methods and techniques of data collection and info generation, that involves large groups of users, who are not organized centrally and generate shared content.” Crowdsourcing has been particularly beneficial for both the collection and generation of spatial data (Papadopoulou & Giaoutzi 2014). As an example, Geo-Wiki is an internet crowd sourcing platform where volunteers improve the quality of land-cover maps using datasets in Google Earth (http://www.geo-wiki.org/). Penguin watch (https://www.penguinwatch.org/) uses citizen science to verify species identities in remote camera footage. Two examples of crowdsourcing platforms are Mechanical Turk (https://www.mturk.com/mturk/welcome) and Tomnod (www.tomnod.com). The latter is specifically for satellite imagery applications and was used recently in the search for missing flight MH370, whereby volunteers were recruited to search satellite imagery for plane debris.

Automated and semi-automated methods

At the time of writing, all studies we could find that use automated techniques to estimate animal numbers by the direct detection of individual animals from RS imagery were proof-of-concept work applied to relatively small areas of usually no more than a few square kilometers, in relatively homogenous environments (e.g. Laliberte & Ripple 2003; Groom et al. 2013; Terletzky 2013; Fretwell, Staniland & Forcada 2014; Table 1). The first two attempts to automate counts of individual animals from 1m resolution satellite imagery were made using a signal model developed for antisubmarine warfare and mine detection (Abileah 2002), and using software developed for medical imaging (ImageTool) (Laliberte and Ripple, 2003). Technologies have emerged in ecology for automated and semi-automated estimation of animal numbers in areas which are difficult to access, where ground counts are challenging, or animals are at low densities over large areas, e.g. emperor penguins (Barber-Meyer, Kooyman & Ponganis 2007), polar bears (Ursus maritimus, LaRue et al. 2015), and southern right whales (Eubalaena australis, Fretwell, Staniland & Forcada 2014). The capability of image processing software to incorporate texture, shape, enumeration, and context has improved object detection and classification (Laliberte & Ripple 2003; Peña-Barragán et al. 2011). Expert judgement can also improve automated techniques by, for example, identifying sources of systematic error (Martin et al. 2012; McBride & Burgman 2012; Burgman 2015).

Image segmentation

Discriminating animals from their environment in imagery depends more on their contrast with the background (Figure 1) than spatial resolution (Laliberte & Ripple 2003). Distinguishing objects of interest from the background in image processing is done by image segmentation. Thresholding is the simplest and most common method, whereby pixels are categorised into multiple features based on an intensity value relative to a threshold value. Thresholding has had some success in small spatial areas, with population estimates usually highly correlated with ground or manual counts (Laliberte & Ripple 2003; Trathan 2004; Fretwell, Staniland & Forcada 2014; Table 1). Trathan (2004) used automated image segmentation and a threshold filter to eliminate background features and classify penguin pixels with aerial imagery to detect macaroni penguins (Eudyptes chrysolophus). The results were compared with both manual counts in the imagery and ground counts at colonies. At two large colonies (~11,000 and ~35,000 birds) the manual counts were 3% and 6% greater than automated counts, and at a smaller colony both methods detected 469 birds. Image segmentation has been shown to outperform other automated methods in some situations, including supervised and unsupervised classification (Fretwell, Staniland & Forcada 2014).

Filtering and image enhancements before processing can enhance images containing small animals that may have been missed due to the small number of pixels they occupy, or by exaggerating differences between a feature and background to improve automated detection of the feature of interest. However, the benefits arising from filtering are situation-dependent as it can lead to misclassification (Laliberte & Ripple 2003), or distort the spectral information of pixels depicting animals (Yang 2012). Objects other than the animals of interest can affect automated counts by inappropriate segmentation of pixels (Laliberte & Ripple 2003).

Supervised and unsupervised classification

Supervised classification is used commonly for animal identification in RS imagery (e.g. Barber-Meyer, Kooyman & Ponganis 2007; Fretwell, Staniland & Forcada 2014; LaRue et al. 2014). Users classify known objects that train the image processing algorithms. The mean and variance of the spectral signatures of the training data are then used to classify the remaining image pixels. LaRue et al (2015) used reflectance values of polar bears (U. maritimus) to classify pixels into bears and background from Rowley Island, Canada. The spectral signatures of bears were not sufficiently different from non-target objects and therefore while all known bears were positively identified, the method also identified thousands of non-target objects as bears. This method did not perform as well as image differencing for polar bear counts. In other studies regression methods have be used to estimate populations by relating clusters of spectrally similar pixels classified as animals, such as emperor penguins, to population size (Barber-Meyer, Kooyman & Ponganis 2007; Fretwell et al. 2012). In this case, individual animals are not detected but pixels containing animals can be counted to estimate populations where large congregations of animals exist in colonies or herds.

Despite its prevalent use, supervised classification has not been as successful at detecting individual animals as other methods (Fretwell, Staniland & Forcada 2014; LaRue et al. 2015; Table 1). The quality of supervised classification depends heavily on the user’s knowledge and ability to correctly identify training data, the distinctiveness of the spectral signatures, and accurately representing the variability in classes representing animals within the training data (Hussain et al. 2013).

Unsupervised classification uses statistical algorithms to group pixels based on spectral information, identifying unique features in a landscape with limited user inputs. Detection probability using this method can be high (mean=80% for livestock, with range 55%-100% across seven images) and comparable with manual counts, but high levels of over-counting have been demonstrated (mean commission error=69%, range 28-98%), thereby overestimating populations. The number of non-detections also increased as the number of animals increased (Terletzky 2013; Terletzky & Ramsey 2016). This method performed significantly better than short time image differencing, described later in this section (Terletzky & Ramsey 2016).

Determining spectral signatures of animal species or specific features

Spectral signatures and profiles can discriminate animal species in satellite imagery. Older studies had little success in determining a unique spectral signature in the visible-near infrared range for deer (Trivedi, Wyatt & Anderson 1982; Wyatt et al. 1985). More recently, spectral or thermal profiles have distinguished large livestock species including sheep (Ovis aries), pigs (Sus scrofa domesticus), horses (Equus caballus) and cows (Bos taurus), and between mammals and landscapes (Bortolot & Prater 2009; Terletzky, Ramsey & Neale 2012; Yang 2012). Spectral separability between mammals and shadows remains poor (Yang 2012). The spectral signature of guano has been used to distinguish seabird colonies from background geology and vegetation (Fretwell et al. 2015). These studies use different source data, methods to detect spectral separability, and procedures to assess spectral overlap between different animal species, and animals and background.

Spectral signatures of species can be used as training data and applied to new imagery to classify pixels into predefined classes (Turner et al. 2003; Bortolot & Prater 2009). Bortolot and Prater (2009) used stepwise discriminant analysis of hyperspectral data to assign pixels to four animal species classes (cattle, horses, sheep and pigs) and generated accurate correlations (>90%) between estimated population sizes and animal surveys, even with relatively coarse image resolution of 2m.

Short time image differencing and change detection

Change-detection methods using multi-temporal imagery have been commonly used to determine land use changes, such as deforestation, due to the consistent, repetitive availability of imagery (Singh 1989). Recently several studies have used these techniques to estimate animal populations. Animals have been detected in small scale studies by the change in spectral reflectance of pixels in two sets of images resulting from animal movements relative to the static background (e.g. Oishi & Matsunaga 2014; Terletzky & Ramsey 2014; LaRue et al. 2015). LaRue et al. 2015 found image differencing detected 87% of known polar bears on Rowley Island in Canada compared with 100% by supervised classification. However, image differencing was more effective than supervised classification at estimating polar bear populations, due to significantly less false positives. Terletzky and Ramsey (2014) estimated livestock numbers using a semi-automated, principal components analysis of two aerial photographs captured on the same day. Polygons representing cattle and horses were generated by using the difference in the first principal component of the images and heuristic thresholding. Using this method, 82% of animals were correctly identified, though mean commission error was high at 53%, attributed to small mismatches in the alignment of separate images, misidentification of shadows, and animals grouped together.

Oishi and Matsunaga (2014) also described automated methods for detecting animal movement through snow from overlapping aerial photographs. Their algorithms for image registration and detection of moving animals successfully detected 5 of 6 cattle and a deer in two UAV images covering approximately 0.6km2 each. A walking human could also be differentiated from among more than 2,100 objects. Misclassification increased with more stringent thresholding, but applying the algorithms reduced the hours required to manually survey animals by over 90%.

Change detection techniques have been used to manually count polar bears (U. maritimus) from high resolution satellite imagery (Stapleton et al. 2014). Objects that could be polar bears were compared to images taken at a different time to eliminate features that remained constant. This method had higher precision (n=94; 95% CI=92-105) than counts from established aerial survey methods (n=102, 95% CI=69-152) (Stapleton et al. 2014). Automated and semi-automated methods, such as those used by LaRue et al. (2015) for image differencing, are recommended for large-scale studies of more than a few square kilometres, largely due to the time required for manual counts.

Change detection in landscape analysis is often applied to low and medium resolution imagery. It has been unsuccessful for very high resolution (VHR) imagery (sub-metre resolution), because of insufficient orthorectification to exactly match two images, leading to mismatches in landscape features which make them appear to be different (LaRue, Stapleton & Anderson 2017). Higher variability in spectral reflectance also complicates discrimination between real changes and background noise (Hussain et al. 2013). Despite the lower resolution requirements, cost may be a constraint because two sets of images are required and should be obtained from the same sensor, at the same time of day, less than a week apart to minimize changes in sun angle and vegetation (Hussain et al. 2013; Terletzky & Ramsey 2014).

Object-based image analysis

Most animal identification studies have used pixel-based approaches, but alternative object-based image analysis (OBIA) methods have been suggested, and have shown improvement over pixel-based approaches in some disciplines (Blaschke 2010). The unit for analysis in OBIA is the object, and neighbouring objects can provide context in spatial relationships, texture, and shape (Blaschke 2010; Hussain et al. 2013). Object-based methods have been mainly used for land cover and land use classification (e.g. Ke, Quackenbush & Im 2010; Peña-Barragán et al. 2011) where they are generally considered superior to traditional pixel-based methods because they reduce spectral overlap between classes, can incorporate expert knowledge, consider both spectral and spatial information, and produce greater classification accuracy (Yang 2012). This superiority may depend on the image resolution and choice of algorithm (Duro, Franklin & Dubé 2012). The time taken to set up and implement OBIA may impede their widespread implementation, mainly as a result of software limitations (Duro, Franklin & Dubé 2012); running OBIA algorithms in eCognition for example can take up to 20 minutes per square kilometre.

Object-based approaches have detected individual animals (Groom et al. 2011; Groom et al. 2013; Chrétien, Théau & Ménard 2016). An automated OBIA using quadtree image segmentation and sequential object brightness thresholding to estimate lesser flamingo (Phoeniconaias minor) abundance demonstrated detection accuracy was >99% when compared with visual counts at a 525ha site, with more than 81,000 individuals identified (Groom et al. 2011). Quadtree is one type of image segmentation option which divides a raster image into square objects based on the relative values of neighbouring pixels. Generally, the process underestimated flamingo counts compared to visual counts. Higher detection rates and lower omission and commission rates were also observed over water than on land because of the greater contrast with background (visual counts as a percentage of automated counts, water: range -0.34% to -5.68%; land: range -0.99% to -21.82%). In this context OBIA was presumed superior to pixel-based approaches because the variability in brightness between adjoining images, and the lower contrast between birds and background on the land than in the water, limited the use of pixel-based approaches across the entire site.

A study comparing pixel and OBIA methods for animal detection, namely migrating mammals in African savannahs, failed to find a difference in performance of the methods for population estimation and pixel identification (Pixel-based: commission error=11-15%, omission error=12%; OBIA: commission error=7-13%, omission error=13-16%; Yang 2012) A second study found OBIA methods using thermal infrared multispectral data to be superior to supervised and unsupervised pixel-based classification, with the pixel-based approaches ineffective at detecting white-tailed deer. OBIA methods had an average detection rate of 50% and were comparable to aerial surveys (Chrétien, Théau & Ménard 2016).

Limitations and constraints of detecting animals from remotely sensed imagery.

Individual animals have been successfully detected from RS data using automated and semi-automated techniques, but to date studies have generally been on small areas of no more than a few square kilometres (e.g. Laliberte & Ripple 2003; Terletzky 2013; McMahon et al. 2014), and/or confined to relatively homogenous environments, where contrast between animals and background is high, for example white Lesser flamingos against water and land (Groom et al. 2011). Accurate animal detection over large areas has only been successful in comparatively homogenous polar regions. Most remote sensing analysis techniques have limitations (Table 2) but algorithms which have been applied successfully over small areas with several species and imagery types, have great future potential (Laliberte & Ripple 2003; Table 1).

Logistical costs of obtaining, processing and analysing RS imagery, technical capabilities and software are still prohibitively expensive for many governments and organisations (Pettorelli et al. 2014), despite declining costs of high resolution imagery and increasing computational power. Factors such as different satellite sensors, different pixel sizes, scales, or numbers of bands and spectral information can affect the classification accuracy and may also prevent images from being accurately overlaid (Hussain et al. 2013). Users should also be aware that different pre-processing methods can give radically different results. For example, image quality and visual interpretability can vary with different image filtering and kernelling methods (Yang 2012). Situations like these may be unavoidable when working over large spatial extents where different image sets are required.

Corroborating image analysis results with ground-based estimates is a key requirement because factors such as atmospheric conditions can influence spectral results and affect outputs (Turner et al. 2003). All methods that detect individual animals from RS imagery suffer from detection errors. For example, animals that cluster or herd in tight groups can be difficult to count (Cunningham, Anderson & Anthony 1996), animals can be obscured by vegetation and young animals can make it difficult to differentiate individuals. There is general recognition that visibility bias can affect outcomes (e.g., Jachmann 2002, McMahon et al. 2014), and work has been done to develop models that will handle imperfect detection in aerial surveys (e.g. Samuel et al. 1987).

Software

GIS software (e.g. QGIS and ArcGIS) is used in most projects that aim to identify objects from RS imagery in ecological and environmental applications, due to the familiarity of users within these disciplines. Other specific remote-sensing software may have more utility and power, such as Erdas Imagine (Hexagon Geospatial) and ENVI (Envivo Inc.) (LaRue et al. 2015). Some studies have combined the use of GIS software with specific image processing software, such as Fretwell et al. (2014) who used ENVI5 and ArcGIS to analyse satellite imagery in the first successful study using RS imagery to count whales. Other studies have borrowed image processing software from biomedical and health disciplines for image analysis, including Laliberte and Ripple (2003; ImageTool) and Groom et al. (2011; Definiens Developer). OBIA analysis is supported in ENVI and eCognition software, and eCognition has been used to count animals in savannahs (Yang 2012). There has been no direct comparison of the efficacy of GIS software and other image processing software for the identification of animals or objects in RS imagery.

Computer pattern recognition, object counting techniques and image processing software are widely used and well developed in many scientific disciplines, including biomedical science and engineering. They may represent valuable and underused alternatives in ecology, and some are now being adapted to identify and count animals (Abileah 2002; Laliberte & Ripple 2003; Groom et al. 2011; Yang 2012; Conn et al. 2014). Image processing software programs available and commonly used in other disciplines include Microsoft Visual Studio, C#, C++. OpenCV, Mathematica, Python, Matlab, ImageJ and Erdas Imagine. The Image Processing Toolkit in Matlab was used for the estimation of penguin population size from aerial photography (Trathan 2004).Otherwise, Matlab has rarely been used in ecology for image processing, despite its wide application in other disciplines, particularly engineering. Descamps et al (2011) developed a birth-and-death algorithm in C++ to automatically count aggregations of birds. ImageJ is open source image processing software that is comparatively easy to use, and many people have developed macros that can be used or modified. (Figure 1).

Future directions

Challenges for the identification and enumeration of individual animals from RS imagery are increasingly being addressed through advances in software, and accessibility to lower cost, higher spectral and spatial resolution images (<1m). This is demonstrated with an increasing number of proof-of-concept studies (e.g. Terletzky & Ramsey 2014; Yang et al. 2014; LaRue et al. 2015), and rapidly expanding image processing technology. Image analysis developments in other disciplines continue unabated, such as microbiology and sub-disciplines in engineering (e.g. Schindelin et al. 2012; Fredborg et al. 2015; Puchkov 2016), and inter-disciplinary collaborations could offer novel opportunities. Detection of different species (e.g. cow versus horse) and in non-homogenous environments remains problematic. Several RS studies used 50cm resolution satellite imagery from common satellites such as Geoeye1, Worldview and Quickbird2 (e.g. Fretwell, Staniland & Forcada 2014; Yang et al. 2014), and other studies have used aerial imagery which is sub-meter resolution (e.g. Groom et al. 2011; McNeill et al. 2011). Higher spectral and spatial resolution imagery will improve our ability to detect small objects such as animals. This is because smaller pixel sizes and a greater number of spectral bands allows greater distinction between animals and background signatures, and provides the potential to automate the identification of spectral thresholds between animals and background. Imagery with higher spectral and spatial resolution are becoming more readily available, but cost can still be prohibitive.

There is a general lack of consistency in performance measures in animal detection literature (Table 1). Applying consistent accuracy measures across ecological studies would aid the future development and application of remote sensing technologies and methods for imagery analysis, such as OBIA and change detection techniques. This would enable comparison, assessment, and evaluation of different methods across a range of ecological applications. This should include the probability of detection, probability of undercounting (omission error) and probability of over-counting animals (commission error) (Terletzky & Ramsey 2016).

Conclusions

Remotely sensed imagery offers exciting new opportunities in ecology, and an increasing number of studies are demonstrating the efficacy of RS technologies for estimating wild and domestic animal populations. The direct detection and counting of individual animals to establish an accurate population measure using automated and semi-automated techniques is still problematic in most situations, particularly in non-homogenous environments, and currently ineffective for most large-scale applications. However, identifying large population fluctuations (e.g. ~1,000’s) in colonies or detecting the presence of animals in remote locations can be achieved with current remote sensing technology (Barber-Meyer, Kooyman & Ponganis 2007) Future developments in the analysis of RS data will improve direct detection capabilities, including the development of algorithms, the crossover of programs from other disciplines, availability, accessibility, cost, and resolution of data. The proliferation of studies assessing direct animal detection, concomitant with advances in technology, reflect the demand and potential for uptake within ecology. Use and application of RS to detect animals will increase when major limitations are overcome, including high errors of omission and commission. In the near future detection of animals in remotely sensed imagery could offer radical changes in the way some animal populations are monitored, offering exciting advances in animal ecology, disease control, conservation and agriculture.

Acknowledgements

The authors would like to thank the Centre of Excellence for Biosecurity Risk Analysis (CEBRA) and the New Zealand Ministry for Primary Industries (MPI) for funding this research. We would also like to thank the three anonymous reviewers whose valuable comments considerably improved our manuscript.

Data Accessibility

This manuscript does not use any data

Author Contributions

TH wrote the first draft of the manuscript. All authors contributed to the ideas, editing and review of subsequent drafts.

References

Abileah, R. (2002) Marine mammal census using space satellite imagery. US Navy Journal of Underwater Acoustics, 52, 709-723.

Aerometrex (2016) Services: Digital Aerial Photography Capture. http://www.aerometrex.com.au/services.html. Accessed 26 July 2017.

Anderson, K. & Gaston, K.J. (2013) Lightweight unmanned aerial vehicles will revolutionize spatial ecology. Frontiers in Ecology and the Environment, 11, 138-146.

Barber-Meyer, S.M., Kooyman, G.L. & Ponganis, P.J. (2007) Estimating the relative abundance of emperor penguins at inaccessible colonies using satellite imagery. Polar Biology, 30, 1565-1570.

Blaschke, T. (2010) Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, 65, 2-16.

Bortolot, Z.J. & Prater, P.E. (2009) A first assessment of the use of high spatial resolution hyperspectral imagery in discriminating among animal species, and between animals and their surroundings. Biosystems Engineering, 102, 379-384.

Burgman, M.A. (2015) Trusting judgements: how to get the best out of experts. Cambridge University Press.

Chrétien, L.P., Théau, J. & Ménard, P. (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.

Conn, P.B., Ver Hoef, J.M., McClintock, B.T., Moreland, E.E., London, J.M., Cameron, M.F., Dahle, S.P. & Boveng, P.L. (2014) Estimating multispecies abundance using automated detection systems: ice‐associated seals in the Bering Sea. Methods in Ecology and Evolution, 5, 1280-1293.

Cunningham, D.J., Anderson, W.H. & Anthony, R.M. (1996) An image-processing program for automated counting. Wildlife Society Bulletin, 24, 345-346.

Descamps, S., Béchet, A., Descombes, X., Arnaud, A. & Zerubia, J. (2011) An automatic counter for aerial images of aggregations of large birds. Bird Study, 58, 302-308.

Deuter, M. (2016) 20 things they don't tell you about UAVs. Aerometrex, From High Above: digital mapping and photogrammetry.

Duro, D.C., Franklin, S.E. & Dubé, M.G. (2012) A comparison of pixel-based and object-based image analysis with selected machine learning algorithms for the classification of agricultural landscapes using SPOT-5 HRG imagery. Remote Sensing of Environment, 118, 259-272.

Fredborg, M., Rosenvinge, F.S., Spillum, E., Kroghsbo, S., Wang, M. & Sondergaard, T.E. (2015) Automated image analysis for quantification of filamentous bacteria. BMC Microbiology, 15, 255.

Fretwell, P.T., LaRue, M.A., Morin, P., Kooyman, G.L., Wienecke, B., Ratcliffe, N., Fox, A.J., Fleming, A.H., Porter, C. & Trathan, P.N. (2012) An emperor penguin population estimate: the first global, synoptic survey of a species from space. PloS one, 7, e33751.

Fretwell, P.T., Phillips, R.A., Brooke, M.d.L., Fleming, A.H. & McArthur, A. (2015) Using the unique spectral signature of guano to identify unknown seabird colonies. Remote Sensing of Environment, 156, 448-456.

Fretwell, P.T., Staniland, I.J. & Forcada, J. (2014) Whales from space: counting southern right whales by satellite. PloS one, 9, e88655.

Groom, G., Krag Petersen, I., Anderson, M.D. & Fox, A.D. (2011) Using object-based analysis of image data to count birds: mapping of Lesser Flamingos at Kamfers Dam, Northern Cape, South Africa. International Journal of Remote Sensing, 32, 4611-4639.

Groom, G., Stjernholm, M., Nielsen, R.D., Fleetwood, A. & Petersen, I.K. (2013) Remote sensing image data and automated analysis to describe marine bird distributions and abundances. Ecological Informatics, 14, 2-8.

Hussain, M., Chen, D., Cheng, A., Wei, H. & Stanley, D. (2013) Change detection from remotely sensed images: From pixel-based to object-based approaches. ISPRS Journal of Photogrammetry and Remote Sensing, 80, 91-106.

Jachmann, H. (2002) Comparison of aerial counts with ground counts for large African herbivores. Journal of Applied Ecology, 39, 841-852.

Kadlec, J.A. & Drury, W.H. (1968) Aerial estimation of the size of gull breeding colonies. The Journal of Wildlife Management, 32, 287-293.

Ke, Y., Quackenbush, L.J. & Im, J. (2010) Synergistic use of QuickBird multispectral imagery and LIDAR data for object-based forest species classification. Remote Sensing of Environment, 114, 1141-1154.

Laliberte, A.S. & Ripple, W.J. (2003) Automated wildlife counts from remotely sensed imagery. Wildlife Society Bulletin, 31, 362-371.

Landinfo Worldwide Mapping LLC (2017) Buying Satellite Imagery: Pricing Information for High Resolution Satellite Imagery. http://www.landinfo.com/satellite-imagery-pricing.html. Accessed 12th April 2017.

LaRue, M., Lynch, H., Lyver, P., Barton, K., Ainley, D., Pollard, A., Fraser, W. & Ballard, G. (2014) A method for estimating colony sizes of Adélie penguins using remote sensing imagery. Polar Biology, 37, 507-517.

LaRue, M.A., Rotella, J.J., Garrott, R.A., Siniff, D.B., Ainley, D.G., Stauffer, G.E., Porter, C.C. & Morin, P.J. (2011) Satellite imagery can be used to detect variation in abundance of Weddell seals (Leptonychotes weddellii) in Erebus Bay, Antarctica. Polar Biology, 34, 1727-1737.

LaRue, M.A., Stapleton, S. & Anderson, M. (2017) Feasibility of using high‐resolution satellite imagery to assess vertebrate wildlife populations. Conservation Biology, 31, 213-220.

LaRue, M.A., Stapleton, S., Porter, C., Atkinson, S., Atwood, T., Dyck, M. & Lecomte, N. (2015) Testing methods for using high‐resolution satellite imagery to monitor polar bear abundance and distribution. Wildlife Society Bulletin, 39, 772-779.

Leonard, R.M. & Fish, E.B. (1974) An aerial photographic technique for censusing lesser sandhill cranes. Wildlife Society Bulletin, 2, 191-195.

Löffler, E. & Margules, C. (1980) Wombats detected from space. Remote Sensing of Environment, 9, 47-56.

Lynch, H.J., White, R., Black, A.D. & Naveen, R. (2012) Detection, differentiation, and abundance estimation of penguin species by high-resolution satellite imagery. Polar Biology, 35, 963-968.

Martin, T.G., Burgman, M.A., Fidler, F., Kuhnert, P.M., LOW‐CHOY, S., McBride, M. & Mengersen, K. (2012) Eliciting expert knowledge in conservation science. Conservation Biology, 26, 29-38.

McBride, M.F. & Burgman, M.A. (2012) What Is Expert Knowledge, How Is Such Knowledge Gathered, and How Do We Use It to Address Questions in Landscape Ecology? Expert Knowledge and Its Application in Landscape Ecology (eds A.H. Perera, C.A. Drew & C.J. Johnson), pp. 11-38. Springer New York, New York, NY.

McMahon, C.R., Howe, H., van den Hoff, J., Alderman, R., Brolsma, H. & Hindell, M.A. (2014) Satellites, the All-Seeing Eyes in the Sky: Counting Elephant Seals from Space. PloS one, 9, e92613.

McNeill, S., Barton, K., Lyver, P. & Pairman, D. (2011) Semi-automated penguin counting from digital aerial photographs. Geoscience and Remote Sensing Symposium (IGARSS), 2011 IEEE International, pp. 4312-4315.

Mejias, L., Duclos, G., Hodgson, A. & Maire, F. (2013) Automated marine mammal detection from aerial imagery. 2013 OCEANS - San Diego, pp. 1-5. San Diego, CA, 2013.

Oishi, Y. & Matsunaga, T. (2014) Support system for surveying moving wild animals in the snow using aerial remote-sensing images. International Journal of Remote Sensing, 35, 1374-1394.

Papadopoulou, C.-A. & Giaoutzi, M. (2014) Crowdsourcing as a Tool for Knowledge Acquisition in Spatial Planning. Future Internet, 6, 109-125.

Peña-Barragán, J.M., Ngugi, M.K., Plant, R.E. & Six, J. (2011) Object-based crop identification using multiple vegetation indices, textural features and crop phenology. Remote Sensing of Environment, 115, 1301-1316.

Pettorelli, N., Laurance, W.F., O'Brien, T.G., Wegmann, M., Nagendra, H. & Turner, W. (2014) Satellite remote sensing for applied ecologists: opportunities and challenges. Journal of Applied Ecology, 51, 839-848.

Puchkov, E. (2016) Image Analysis in Microbiology: A Review.

Robinson, T.P., Wint, G.W., Conchedda, G., Van Boeckel, T.P., Ercoli, V., Palamara, E., Cinardi, G., D'Aietti, L., Hay, S.I. & Gilbert, M. (2014) Mapping the Global Distribution of Livestock. PloS one, 9, e96084.

Samuel, M.D., Garton, E.O., Schlegel, M.W. & R.G., C. (1987) Visibility bias during aerial surveys of elk in north central Idaho. Journal of Wildlife Management, 51, 622–630.

Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., Preibisch, S., Rueden, C., Saalfeld, S., Schmid, B., Tinevez, J.-Y., White, D.J., Hartenstein, V., Eliceiri, K., Tomancak, P. & Cardona, A. (2012) Fiji - an Open Source platform for biological image analysis. Nature methods, 9, 10.1038/nmeth.2019.

Singh, A. (1989) Review article digital change detection techniques using remotely-sensed data. International Journal of Remote Sensing, 10, 989-1003.

Stapleton, S., LaRue, M., Lecomte, N., Atkinson, S., Garshelis, D., Porter, C. & Atwood, T. (2014) Polar Bears from Space: Assessing Satellite Imagery as a Tool to Track Arctic Wildlife. PloS one, 9, e101513.

Terletzky, P. & Ramsey, R.D. (2014) A Semi-Automated Single Day Image Differencing Technique to Identify Animals in Aerial Imagery. PloS one, 9, e85239.

Terletzky, P., Ramsey, R.D. & Neale, C.M. (2012) Spectral Characteristics of Domestic and Wild Mammals. GIScience & Remote Sensing, 49, 597-608.

Terletzky, P.A. (2013) Utilizing Remote Sensing and Geospatial Techniques to Determine Detection Probabilities of Large Mammals. PhD Dissertation. PhD, Utah State University.

Terletzky, P.A. & Ramsey, R.D. (2016) Comparison of three techniques to identify and count individual animals in aerial imagery. Journal of Signal and Information Processing, 7, 123.

Trathan, P.N. (2004) Image analysis of color aerial photography to estimate penguin population size. Wildlife Society Bulletin, 32, 332-343.

Trivedi, M.M., Wyatt, C.L. & Anderson, D.R. (1982) A multispectral approach to remote detection of deer. Photogrammetric Engineering and Remote Sensing, 48, 1879-1889.

Turner, W., Spector, S., Gardiner, N., Fladeland, M., Sterling, E. & Steininger, M. (2003) Remote sensing for biodiversity science and conservation. Trends in ecology & evolution, 18, 306-314.

Velasco, M. (2009) A Quickbird’s eye view on marmots. Masters Thesis. MSc Thesis, International Institute for Geo-information science and Earth Observation.

Williams, M. & Dowdeswell, J.A. (1998) Mapping seabird nesting habitats in Franz Josef Land, Russian High Arctic, using digital Landsat Thematic Mapper imagery. Polar research, 17, 15-30.

Wyatt, C.L., Anderson, D., Harshbarger, R. & Trivedi, M. (1985) Deer census using a multispectral linear array instrument. Proceedings of the 18th International Symposium on Remote Sensing of Environment, Paris, France, pp. 1475-1488.

Yang, Z. (2012) Evaluating high resolution GeoEye-1 satellite imagery for mapping wildlife in open savannahs. PhD Thesis. Enschede, The Netherlands University of Twente.

Yang, Z., Wang, T., Skidmore, A.K., de Leeuw, J., Said, M.Y. & Freer, J. (2014) Spotting East African mammals in open savannah from Space. PloS one, 9, e115989.

Table 1: Peer reviewed journal articles applying automated and semi-automated techniques to RS imagery to directly detect animals.

Reference

Target species

Location

Method

Remote sensing imagery type

Image resolution

Population size

Geographic extent (where not explicitly specified, the number of images is given)

Measure of accuracy

(Laliberte & Ripple 2003)

Caribou (Rangifer tarandus) and cattle (Bos taurus)

Caribou and cattle, USA

Thresholding

Caribou: Black-and-white aerial imagery;

Cattle: Very high Resolution (VHR) satellite imagery

Caribou: taken at 500 m altitude with a 22.86-cm mapping camera

Cattle: 1m

Caribou: 60-406

Cattle: >50

1 image for each species

Mean count error for aerial photos of caribou 10.2%

(Trathan 2004)

Macaroni penguin (Eudyptes chrysolophus)

Bird Island, South

Georgia

Image

Processing Toolbox (Matlab); image segmentation, thresholding

Aerial photography

0.05 m ground resolution

3 penguin colonies with >50,000 birds

Island size 5.0km x 0.8km

Automated image-analysis highly correlated (R2 = 0.95-0.98) with ground counts. Manual counts 3-6% higher than automated counts

(Barber-Meyer, Kooyman & Ponganis 2007)

Emperor penguins (Aptenodytes forsteri)

Antarctica

Supervised classification

VHR satellite imagery

0.6 m panchromatic imagery

Predict

relative abundances for two categories: <3,000 or >5,000 adult birds;

12 images covering seven penguin colonies located up to 600km apart

Absolute percent deviation between

predicted adult birds and known adult birds averaged 53% (SE = 15) and ranged from <1 to 128%

(Groom et al. 2011)

Lesser Flamingo (Phoeniconaias minor)

Kamfers Dam, South Africa

OBIA: quadtree image segmentation and sequential object brightness

thresholding

Aerial imagery,

0.08 m ground resolution

81,664 flamingos;

31 images

99% detection compared to human visual interpretation. Under-estimation by the object-based method of less than 0.5%

(McNeill et al. 2011)

Adélie penguins (Pygoscelis adeliae)

Antarctica

Semi-automated software written in Matlab; linear discriminant analysis to separate the background; morphological image processing operators to select breeding penguins

Aerial imagery,

0.5 m or less resolution

Not available

10 images

Detection generally above 90-95%. Omission error generally less than 10%. Commission error generally below 25% but as high as 65% in one image

(Descamps et al. 2011)

Greater Flamingo (Phoenicopterus roseus)

France, Turkey and Mauritania

Stochastic approach based on object processes; a birth-and-death algorithm

Aerial imagery

~2000 to greater than 10,000

Two islands

Omission error <12.9% of birds; Commission error from 0 - 15%. The overall error was 2 - 5%.

(Mejias et al. 2013)

Dugongs (Dugong dugon)

Shark Bay, Western Australia

Morphological-based detection; Shape profiling on saturation channel

Aerial imagery (UAV)

100 images taken at 1000ft

Recall values of 48.57% and 51.4%, precision values of 4.01% and 4.97%, and commission error of 43 and 38 objects per image for the two algorithms respectively.

(Groom et al. 2013)

19 marine bird species/ species groups

Two offshore wind farms, Irish Sea

OBIA

Aerial imagery

0.03-0.04 m

~260 birds

Total area ~670 km2; 18 image frames

The overall success of detection is 92.5%

(Yang et al. 2014)

Wildebeest (Connochaetes taurinus), zebra (Equus quagga burchellii), and gazelle (Eudorcas thomsonii)

Maasai Mara National Reserve, Kenya 

Artificial neural network applied to spectral reflectance

VHR satellite imagery

0.5 m

Hundreds of animals

Two pilot study areas 1km x 1km each;

Average count error of 8.2%, omission error of 6.6% and commission error of 13.7%

(Terletzky & Ramsey 2014)

Cattle and horses (Equus caballus)

Utah, USA

Short-time interval image differencing using principal components analysis

Aerial imagery

0.25 m

158 animals

Eight pastures/paddocks

82% of animals correctly identified, mean commission error was 53%, and mean omission error was 18%.

(Fretwell, Staniland & Forcada 2014)

Southern Right Whales (Eubalaena australis)

Golfo Nuevo, Penı´nsula Valde´s in Argentina

Various; thresholding the best performing

VHR satellite imagery

0.5 m

<100 individuals

113km2

Thresholding: found 84.6% of all manually digitized whales and 89% of objects manually classed as probable whales, with 23.7% commission error

Supervised classification: no meaningful results

(LaRue et al. 2015)

Polar bears (Ursus maritimus)

Rowley Island, Canada

Supervised spectral classification; image differencing

VHR satellite imagery

0.5 - 0.65 m

94 bears

1,100km2

Supervised classification detected all known bears but thousands of false positives (commission errors); Automated image differencing correctly identified 87% of bear locations.

Table 2: Advantages and limitations of methods used for detecting animals in RS imagery.

Method

Recommended data attributes

Advantages and reliability

Challenges and limitations

Manual Counting

High spatial resolution

· Has been used extensively for decades

· No high-level technological requirements for image processing

· Can use most types of data that has adequate resolution with respect to animal size

· Small areas have high correlations with traditional survey techniques

· Observer confidence and experience can impact counts

· Costly for labour

· Time consuming

· Subjective and non-repeatable

· Accuracy issues over large areas

· Often underestimates populations

Image Segmentation

High spatial resolution

· Often outperforms other automated techniques

· One of the simplest automated techniques

· Less opportunities for user error, though users have to identify thresholds

· Data is relatively cheap, using panchromatic images

· Potentially high misclassification

· Need to experiment with different thresholds

· Image noise (e.g. shadows, reflections, illumination) and spectral variability of target classes can affect segmentation

· Binary classification of pixels into background and foreground, but variability means that pixel values can overlap

Supervised Classification

High spectral and spatial resolution

· A common approach so methodology well documented

· Potentially more accurate than unsupervised classification if users have good knowledge of environment

· Reliability based on users ability to identify animal pixels in training data

· Full spectral variability of animal pixels needs to be represented, and distinctive compared to non-animal pixels

· Generally not as accurate as other automated techniques

Unsupervised Classification

High spectral and spatial resolution

· Fully automated without need for user to identify spectral classes; limits user error

· Can be used in combination with supervised classification

· Can generate many classes and may require user to combine or restrict the number of classes identified

· Users need to identify which class/es are animals

· Demonstrated high levels of over-counting

Spectral and Thermal Separability

Hyperspectral data

· High correlations (>90%) with hyperspectral data

· Can use with relatively coarse spatial resolution

· Consistency of spectral distinction of pixels can vary between species, and between animals and background

· Ideally requires expensive hyperspectral data

· Only proof-of-concept studies conducted on small areas

Short time image differencing, change detection or image differencing

Multi-temporal imagery with high spatial resolution

· Does not require high resolution data

· Can be used manually or with automated and semi-automated techniques to detect pixel changes

· Doesn’t require user knowledge of the area

· High correlations demonstrated in proof-of-concept studies

· Requires two separate imagery sets

· Requires short time intervals between imagery; therefore generally aerial imagery

· High spectral resolution can affect geo-matching, causing small mismatches in overlaid images and commission errors

· Very new methodology for animal detection

· Only proof-of-concept studies conducted on small areas

Object-based image analysis

High spectral and spatial resolution

· Considered more accurate than pixel based approaches

· Can incorporate expert knowledge and consider both spectral and spatial information

· Neighbouring objects can provide contextual information to improve detection

· Reduces spectral overlap of classes

· Objects are identified based on a pixel groups with similarities – e.g. size, shape, spectral properties

· Requires expensive software and substantial computer and processing power

· Complex method potentially requiring specialist knowledge

· Used mainly for land cover and land use, but also features such as buildings

· Potentially slow processing speeds especially for large areas

· Requires user knowledge to group objects into relevant classes

33

a) 1

2

3

4

5

6

7

8

b) 9

10

11

12

13

14

15

16

a) 1

2

3

4

5

6

7

8

b) 9

10

11

12

13

14

15

16

a)

b)

c)

d)

Figure 1: Particle Tracker (Sbalzarini & Koumoutsakos 2005) in ImageJ, designed for video imaging in cell biology, to detect sheep; a) original image of farming area in NZ at 0.4m resolution; b) particle tracker program for automated detection of ‘particles’ (red circles) to detect sheep; c) a representative area of peri-urban small scale farming in NZ at 0.4m resolution; d) Particle tracker to detect sheep. This heterogeneous landscape has less success with many false detections. Images from http://data.linz.govt.nz