Prying into the intimate secrets of animal lives; software beyond hardware for comprehensive annotation in ‘Daily Diary’ tags
© Walker et al. 2015
Received: 22 January 2015
Accepted: 6 September 2015
Published: 21 September 2015
Smart tags attached to freely-roaming animals recording multiple parameters at infra-second rates are becoming commonplace, and are transforming our understanding of the way wild animals behave. Interpretation of such data is complex and currently limits the ability of biologists to realise the value of their recorded information.
This work presents Framework4, an all-encompassing software suite which operates on smart sensor data to determine the 4 key elements considered pivotal for movement analysis from such tags (Endangered Species Res 4: 123-37, 2008). These are; animal trajectory, behaviour, energy expenditure and quantification of the environment in which the animal moves. The program transforms smart sensor data into dead-reckoned movements, template-matched behaviours, dynamic body acceleration-derived energetics and position-linked environmental data before outputting it all into a single file. Biologists are thus left with a single data set where animal actions and environmental conditions can be linked across time and space.
Framework4 is a user-friendly software that assists biologists in elucidating 4 key aspects of wild animal ecology using data derived from tags with multiple sensors recording at high rates. Its use should enhance the ability of biologists to derive meaningful data rapidly from complex data.
The development of hardware that can be attached to animals during their everyday life  has revolutionized our understanding of the biology of wild animals. Indeed, this approach has allowed researchers to look at everything from the behaviour of whales chasing prey at depths of over 1 km underwater  to the physiology of geese migrating over the Himalayas . A common feature facilitating these sorts of projects has been the increase in numbers and types of sensors used in smart animal tags, as well as increases in the frequency with which they can be sampled and concomitant increases in memory capacity. Thus, our ability to answer critical questions in biology appears to have been driven to a large extent by advances in technology . These advances in methodology come under two broad areas. One relates to methods that allow tags to be physically attached to animals for increasing lengths of time (e.g. [19, 32, 46]) while minimizing animal detriment [37, 38] while the other relates to the physical production of the complex solid-state units in smart tags (e.g. ).
The excitement at the potential inherent in sophisticated animal tags has, however, been tempered by a new limiting factor. This is a methodology to deal with the problem of the analysis of the high resolution, multiple channel (and therefore multi-dimensional) data acquired by the tags – in short, software (cf. ). To be most useful to the community, software to help in the analysis of smart tag-acquired data needs to be able to deal with large quantities of multiple sensor data and, ideally, should be able to merge different derived analytical outputs together into one output file so that various elements derived from the primary data can be interrelated. Currently, the smart tag community has witnessed a number of software innovations, most of which are concerned with determination of behaviour, i.e. from dive profile data  or, most notably, from dual-axial acceleration data [33, 41]. Analysis of acceleration data is particularly welcome because inspection of raw acceleration values over time to derive behaviours is not particularly intuitive . Thus, solutions for this have involved a suite of different approaches including cluster analysis , support vector machine classification models  and artificial neural networks .
In 2008, Wilson et al.  put forward a concept for a particular sensor configuration within a tag that they called the ‘Daily Diary’ (DD), where analysis yielded value well beyond the simple mathematical sum of its individual parts . Specifically, the suggestion advocated the combination of tri-axial accelerometers, tri-axial magnetometers, and pressure and speed transducers together with environmental sensors such as temperature, light and humidity. The theory was that this constellation of sensors, sampled at infra-second rates, would allow tag users to be able to derive four key elements of animal ecology seamlessly. These are: (1) animal trajectory, and therefore position  (2) animal behaviour (Shepard, ) (3) energy expenditure  and (4) the environmental conditions to which the tag carrier is exposed . Although this original work did indicate avenues by which these elements might be achieved, there was no specific suggestion of software that might actually do this. In short, currently, although some software is available to help determine some aspects of that advocated by the DD concept (e.g. ), there is nothing that binds the concepts together.
This paper describes the structure and functioning of a new software package (Framework4) that allows the users of smart tags to calculate all four key elements advocated by the DD and then to bind them together into one single output file so that workers can subsequently geo-reference behaviours, energy expenditures and environmental parameters to gain a more holistic picture of how animals react to and within their environment. Specifically, we introduce Framework4, a user-friendly turnkey solution for the analysis of smart sensor data. We demonstrate our software on data recorded using a DD, however, the software can be applied to any data formatted to our input specifications (see Section 2.1). Using our system, we can obtain seamless animal behaviour, animal trajectory, energy expenditure, and environmental conditions, all within one application and export them in one merged data file. Our solution requires no knowledge of the underlying processes utilised in the software, or any advanced mathematical and computing skill sets. Our application has been produced with the end-user in mind, utilising wizards and graphical user interfaces wherever possible. We hope this software will assist in the understanding of wild animal ecology, providing new insights as a result of advanced computing knowledge.
Framework4 has been implemented through a five year long collaboration with the Bioscience and Computer Science departments at Swansea University. The result is a software package for the Microsoft Windows operating system for analysing smart sensor data. We create a desktop application as it allows us to handle large data files effectively (tested on over 5 million data samples) on standard computing equipment. Alternatively, utilising a web application would allow access from anywhere on any platform with an active internet connection and a web browser, however, this entails long waiting times while data sets upload, resulting in a reduced ability to handle larger files. Utilising a desktop application, we directly communicate with the CPU for efficient data handling, and the GPU for visualisation purposes. Our tool can be used in the field, in remote areas during deployment where there is often no internet access. Each of the software features and the underlying methods by which they operate is now discussed. Design choices have been made at each stage in order to assist the user in their tasks, for example, making use of wizards to lead the user through various processing tasks.
The software supports two file formats; comma separated values (CSV) and tab delimited formats. These are two of the standard file formats for storing tabular data in text format and are common outputs from commercial smart sensor tags. The DD exports its data in a binary format which gets segmented into multiple files and converted to tabular delimitated format post-deployment. Framework4 loads and operates on the individual tabulated files.
We incorporate an import wizard in the application to import data files. Here, the user can specify the names and data types of each data attribute, which are used as a reference to them throughout the application to allow clear identification.
Derivation of animal trajectory by dead-reckoning
The way in which animals use the environment is fundamental to understanding their behavioural ecology  and, as such, many different systems have been developed to examine animal movements (see e.g. ). GPS has been popular for animal movement. However, fine infra-second scale behaviour cannot be obtained . A relatively recent addition to the field which can achieve this is dead-reckoning , this has particular value in not being dependent on transmission technology. However location cannot be obtained accurately from inertial sensors due to error accumulation. What we achieve in this software is a framework that allows researchers to experiment with location fixing data channels combined with motion channels to provide corrected dead reckoning. It is not the goal of this paper to validate one method over another see Bidder et al. , but rather to provide the framework where animal paths can be integrated from one or a mix of data channels.
Dead-reckoning operates on the basis that the position of an animal at any time ‘t’ can be derived knowing the position of the animal at a previous time ‘t-1’ and the distance and heading taken between the two time intervals. Dead-reckoning has received little interest until now, partially because early systems for dead-reckoning were crude [20, 47, 50] but with the development of sensors and techniques, headings can now be computed to within 1° utilising accelerometer and magnetometer sensors [9, 48] by measuring the earths’ magnetic field.
The earth’s magnetic field is constructed of field lines approximating a magnetic dipole. Each field line originates at a point near the magnetic South Pole and terminates at a point near the magnetic North Pole. Measuring the strength and direction of the field lines using a tri-axial magnetometer can obtain a relative compass heading in respect to magnetic north. The relationship between magnetic north and geographic north is defined by a declination angle which varies across the earth’s surface and with time. The angle of declination can be obtained from a reference table provided by the National Geophysical Data Center (http://www.ngdc.noaa.gov/geomag-web/). Applying the declination angle to the magnetic north heading obtains a geographic heading.
Many algorithms have been introduced to this end for dead-reckoning (e.g. [21, 25, 45]) to provide new insights into animal movement. These methods have been made available to the research community through statistical software packages. Narazaki and Shiomi 2010  introduced ‘ThreeD_path’, a library for the Matlab statistical package based on the dead-reckoning algorithm by Johnson and Tyack  for reconstructing 3-D paths. More recently, Battaile  created the ‘TrackReconstruction’ R package to perform dead-reckoning and enable visualization of the derived trajectories. Both of these packages hold value for those with experience with the associated statistical programming languages, allowing direct manipulation of the methods used and data supplied. However, for those without any background with these applications, the learning curve can be steep, and may appear non-trivial to those with limited experience. We provide a user-friendly protocol which requires no programming experience, allowing the user to see and access the underlying derived data at each step of the dead-reckoning procedure to provide data transparency. Furthermore, the dead-reckoning aspect is tied in closely with behaviour analysis functionality which is not offered in any existing tools.
Each of the components of the dead-reckoning wizard are modular so that the user can select which analysis steps are required. The steps are; (i) GPS import for synchronising time-stamped GPS (or similar) data with the original data file (ii) alignment correction for the accelerometer and magnetometer coordinate frames (iii) heading derivation from the magnetometer channels (iv) speed derivation via a proxy derived from acceleration for obtaining an estimate of the speed, or constant speed options (vi) dead-reckoning to combine the heading and speed information to derive an animal trajectory and (vii) dead-reckoning correction using external positioning information to eliminate drift in the final dead-reckoned path. We refer the reader to Bidder et al.  for the full methods associated to each component of the process. We now briefly introduce these in turn, detailing the software front-end.
It is often the case that positional information is recorded from separate data sources to the rest of the data. For this reason we can import positional data (e.g. GPS) from separate data files and merge it back into the complete data set, although both are assumed time-synchronised. The same import wizard is used as in the file importer to import the geo-referenced data. The user imports the data, then selects the relevant time fields in both the data sets (i.e. day, hours, minutes, seconds, and milliseconds). The wizard matches the time index columns and appends the additional data fields wherever a matching data item with the same time stamp is found. Where no aligning row exists, null values are used, which are ignored during the dead-reckoning procedure.
Deriving the heading necessitates that the data attributes be sanitised via; magnetometer calibration, pitch and roll computation, coordinate frame adjustments, and finally the heading derivation. These processes are executed in the background across separate pages of the wizard and the user is exposed to the parameters and results which allow them to select the correct attributes and associated settings for each stage in the wizard. For clarity we now briefly introduce the methods for each process.
The measurements obtained by commercial magnetometer sensors are corrupted by several errors. Proper calibration of the magnetometer is required to obtain high accuracy heading measurements. Inconsistencies are usually introduced by instrumental errors, such as scale factors, non-orthogonality between axes, offsets and magnetic deviations caused by perturbations and interference with the magnetic field lines.
Magnetic measurements are subject to sources of error primarily caused by hard iron and soft iron deposits acting on the magnetic field. Rotating a magnetometer around 360 degrees in all orientations under no sources of error should produce a perfect sphere centred on the origin. Hard iron effects are created by objects which produce a magnetic field with a constant bias in the output, resulting in a sphere displaced from the origin. Soft iron deposits are caused by ferrous materials which are more permissive to the magnetic field influencing the magnetic field as it passes through, via distortion or stretching. This distorts the sphere into an ellipse as hard iron errors are independent of the orientation of the device and can therefore occur across the sphere. Hard iron distortions are caused by metals such as nickel and iron and commonly have a much larger contribution towards the total error .
We utilise the state-of-the-art error model presented by Renaudin et al.  consisting of an ellipsoid-fitting algorithm based on an adaptive least squares estimator which calibrates the magnetometer readings for both instrumental errors and magnetic deviations. Prior to deployment the user is required to reposition the device at a number of orientations to obtain a complete range of values which can then be fitted to an ellipse and reformed to a uniform sphere. The user selects the magnetometer data channels which are then used to compute the correction matrix. The user can view and adjust the given matrix which can be exported and applied to other data sets from the same deployment.
Pitch and roll computation
In order to determine the heading of a device affixed to an animal, the magnetometer should ideally have an x y plane that is parallel to the earth’s surface, something that is not always possible given the mounting position and uncertainty of animal behaviour. Tilt correction for pitch and roll is computed using the static acceleration derived by passing a windowed moving average over the acceleration axes . Shepard et al.  provide guidelines on selecting an appropriate window size. From the static acceleration the pitch and roll can be computed.
The user selects the accelerometer attributes, along with the window size to use for deriving the pitch and roll. A time-series graph shows a preview of each of the derived angles. The nature of the software allows the user to experiment with different window sizes to see the direct impact of different parameters on the resulting pitch and roll before continuing.
Coordinate frame adjustment
Device attitude via pitch and roll can be used to rotate the magnetometer measurements to bring them back level with the earth’s surface. Within this step, the user selects the pitch and roll channels. This allows the user to select non-derived values, for example pitch and roll, from a gyroscope for this purpose otherwise the pitch and roll from Section 220.127.116.11 should be selected. The user clicks compute and can preview the rotated magnetometer channels in a time-series graph view.
In this step the user selects the adjusted magnetometer components and clicks compute. A preview of the derived heading is displayed for verification.
For volant to correspond to the speed of the animal, invalidating VeDBA in this context. Until a satisfactory measure of speed is derived (such as a pitot tube), we suggest using the constant speed option, and correct this later using positional information (like GPS).
Dead-reckoning combines speed and heading to compute a trajectory for the given data. There are a number of parameters which must be defined first by the user. These are; (i) an initial start position defining the point where the path starts (if GPS data is given, the start coordinates are taken from this), (ii) a number of speed parameters for the VeDBA threshold and speed coefficients (see Bidder et al. 2012 [3–5]) and (iii) a heading offset corresponding to the declination angle obtained from the NGDC website (previous). The computed path is shown alongside in a map below the parameters so that the user can interactively adjust settings and see the result on the generated path.
The complete set of latitude and longitude points defines the trajectory of the animal movement across the earth’s surface. In future work we wish to also derive vertical movement using the pressure sensor. The trajectory columns are appended to the data set once ‘next’ is pressed.
Dead-reckoning with position correction
The dead-reckoning solution can be subject to cumulative errors from the estimates of derived heading and speed. Even small, but systematic, errors in these channels will accumulate over time and thus can increase the resulting error correspondingly.
To overcome these problems, the program utilises a dead-reckoning correction algorithm to correct the heading and speed of the obtained dead-reckoned trajectory using aligned positional trajectory information. This resets the accumulated error at each given position.
Bidder et al.  correct the heading and speed by adjusting the length and heading of the dead-reckoned path until they align to the same positions along each position fix.
Each iteration of the formula makes the path adhere more tightly to the ground truth fixes as the heading and speeds are adjusted. We allow the user to repeat the adjustment process a set amount of times or continue until the speed and heading adjustments are under a specified threshold. Once finalised the user clicks ‘finish’ where the corrected latitude and longitude components are appended to the data set.
Derivation of animal behaviour
Extracting animal behaviour from raw sensed data (most commonly tri-axial accelerometer channels) is a time-consuming and cognitively demanding process for human analysis. Machine learning processes, namely classification, can be applied to identify and label behaviour by executing algorithms which learn from data to discover previously unknown properties . The learning aspect is typically split into two categories; supervised, and unsupervised learning. Supervised learning algorithms are trained on labelled data to generalize relationships between input and output samples. Conversely, unsupervised learning algorithms operate on unlabelled data to find previously unknown structure, and domain knowledge can then be applied to match found structure to a behaviour.
Supervised techniques build a model from labelled data which generates predictions in response to new data. Traditionally, K-nearest neighbour (K-NN), support vector machines (SVM), and random forests have all been applied to accelerometer data. K-NN is a ‘lazy’ learning method which predicts class membership based on the k closest training examples in the feature space (e.g. ). The SVM algorithm finds a hyperplane which separates the feature space into the classes defined in the training set. Unseen data is assigned to a class based on the hyperplane region under which it falls (e.g. ). Random forests are the current state-of-the-art classification method in the data-mining community (e.g. ). Random forests construct many decision trees, each modelling the training set, with each tree voting for the resulting classification. A data item is assigned to the class with the most votes.
Framework4’s approach to behavioural identification
In this section we introduce the behaviour identification functionality of Framework4. Unlike the previous methods, we consider classification as an extended form of search . Providing just one positive example of a behaviour, the user searches for matching behaviours and sorts them into the correct classification group. Complicated parameters are avoided by utilising interactive visual interfaces which draw from the domain expertise when selecting matching behaviours in the input signal. A feedback loop is incorporated such that the precision and recall can be boosted by applying the user’s subject knowledge. These features overcome the disadvantages of machine learning and provide a working solution that can cope with large complex data sets, a vital element given increasing sampling rates (e.g. 10 channels, each at 40 Hz).
The result is a system which supports the manual labelling of animal behaviour complimented with a user-driven approach for the semi-automatic classification of animal behaviour, requiring one instance of behaviour for the matching process to take place.
The components of our system and each of the processes towards classifying animal behaviour are split into five steps; (1) the user must select a behaviour to find in the data set, (2) matching is performed to find similar regions across the series, (3) A classification wizard allows the user to apply their knowledge to extract matches, (4) Extracted behaviours are presented to the user, (5) The user can improve the accuracy of their results by applying a feedback loop, (6) The classified results are displayed.
The first step in our system is for the user to select a behaviour to classify in the data. There are two methods for this in the application. Firstly, query-by-example, and secondly selecting previously saved behaviour instances from the template database.
Query-by-example allows the direct selection of animal behaviour by drawing a window across the time-series encapsulating a subset of data exemplifying the behaviour the user wishes to search for. After selection, a dialog is then displayed where the user can select the data attributes to utilize for the template. Any data attribute can be used for searching throughout the system (not just the accelerometer component). For example, the magnetometer attribute is useful for finding thermalling cycles in condors, while the pressure component can indicate diving cycles in aquatic species.
Behaviour templates used in the system are stored in a database back-end for future use. This is particularly useful because assigned behaviours can be used to search other files. The database can be set up to deal with behaviour templates assigned to classes of animals. The user can query for all patterns present for a specific animal or select an existing behaviour template previously saved in the database by navigating to the animal of interest and then selecting the appropriate behaviour template.
The signal may be resampled to capture events at different frequencies as some behaviours occur at different speeds, for example running. To capture these events independently of the time duration, we can store and search for the signal at different time-intervals using resampling. Re-sampling is used by specifying an irrational factor consisting of an interpolation factor (amount to up-sample by) and a decimation factor (amount to down-sample by)e.g. resampling by 1 / 2 will half the sampling rate, while resampling by 2 / 1 will double the duration.
Supervised machine learning techniques rely on large bodies of labelled data. Such extensively labelled data sets do not exist in our application area. An alternative solution is to consider classification as an extended form of search. That is, to search for matching behaviours, then sort matches into the correct classification group, or reject them. Template matching is a process for determining the presence of a known waveform in a larger dataset. In essence, this works by sliding the specified template across the data set, computing the similarity of the template at each position in the data series corresponding to the concordance in fit between template and the sample at the position. The result is a similarity value at each position in the time-series. This allows the user to select a single positive example of a behaviour and search for all occurrences of it in the data.
A distance measure is used to determine a quantitative value corresponding to similarity or dissimilarity between time-series. Correlation is the optimal technique for detecting a known waveform in random noise . In signal processing, it is well known that Correlation has a linear complexity frequency space. We utilise correlation and a new fast normalized cross-correlation method for template matching in order to obtain results within real time (seconds) which maintains an interactive implementation. Standard cross-correlation performs matching, taking into account amplitude information, while normalized cross-correlation normalizes the template signal and current area under the template such that amplitude shifts are not taken into account. This is important when performing matching over regions where the represented waveform may be present at different orientations. Re-sampling the signal allows us to introduce time axis distortion to extract behaviours occurring at different durations. Walker et al.  specify the exact nature of the algorithm used behind this aspect of Framework4. It is beyond the scope of this paper to discuss the exact methods used, but rather the software functionality available to make them accessible to the research community.
Once the template matching algorithm has been executed, the user is presented with the pattern-matching results in the classification wizard. The classification wizard is used to guide the user through refining a similarity threshold to verify matched signals. The user interactively modifies the threshold value which corresponds to the similarity of extracted matches. Matches are depicted to the user in an intuitive format with interaction to modify the result set according to the user’s domain knowledge.
The classification wizard features two views. On the left (Fig. 6(b)) are visualisations to show where matches occur in the data series, while on the right (Fig. 6(c)), all of the extracted matches are overlaid on top of each other to show the variance between matches. The visualisations are updated as the threshold value is refined by adjusting a slider corresponding to the threshold percentage (Fig. 6(a)).
The positions where matches occur in the series are depicted using three graphical views (Fig. 6(b)), all of which are aligned with a time-series graph of the data series (Fig. 6(b1)). The confidence of a match visualization (Fig. 6(b2)) depicts a heatmap showing an overview of the pattern matching results to encode where high (blue) and low (yellow) similarity matches occur in the data series. The extracted matches view (Fig. 6(b3)) depicts where the extracted matches occur in the series and is updated as the similarity threshold is adjusted. Finally, a distribution of extracted matches (Fig. 6(b4)) utilises a histogram to show the number of matches at each position. The user may refine the result set by rejecting results by clipping rectangular regions of matching results from the data series to reject. This allows the user to reject results based on their knowledge of where they expect results to be present in the data series and the temporal trends expected.
All of the extracted matched signals are overlaid in a stacked time-series graph format, one graph for each data attribute of the pattern (Fig. 6(c)). The user can gain an overview of the general shape of the extracted signals from the graphs. This allows the verification of the shape of extracted matches as most outliers stand out immediately, not fitting into the general shape of the extracted results. The same colour-encoding scheme as the position of matches is used to encode the strength of a match. As the user adjusts the threshold, results are added or removed. The user can directly see the cause and effect of modifying the threshold on the general shape of matched signals in comparison to the template signal overlaid in red. Results can be rejected in this view by manual selection of lines on the time-series graphs. All results falling within the selection are removed from the result set. This allows the user to filter results that should not be present manually. The user continues adjusting the similarity threshold and rejecting results in the synchronised graphical views until they are satisfied with the results being extracted. The user clicks ‘finish’ and the wizard closes.
After the user has concluded with a good threshold value, the results are extracted and added to the results widget in the bottom left of Fig. 6. The user can further inspect the results using two views. Firstly, the sparkline (embedded time-series) display, which puts the classifications in a table format, with each row corresponding to an identified instance of a behaviour visualized using a sparkline. The user can accept or reject results by selecting rows. Secondly, the overlaid plot view overlays the classified instances in a time-series graph. The user can accept or reject results by selection on the time-series. The overlaid plot view is useful where the signals shape is similar amongst results. Conversely, the sparkline display is useful where the behaviour signal varies. Matches displayed in the results view are also shown in the data view overlaid on top of the time-series graph in grey.
Improving precision and recall
The variability and inconsistency of animal behaviour makes the automatic labelling of behaviour a challenging task. It is widely accepted in the machine learning community that achieving 100 % precision and recall is a difficult, if not impossible, task. From a movement ecology viewpoint, we aspire to achieve a close to perfect labelling of behaviour. We incorporate a feedback loop which draws from domain expertise to enhance results. Firstly, the user can provide secondary examples of a behaviour to find more behaviour instances. Secondly, the user can directly manipulate the result set to accept and reject matches. Finally, the user can manually classify behaviour.
Where the user believes the number of found instances to be low, boosting can be used to retrieve more instances. More examples of a behaviour are utilised in the template for searching. This, in effect, widens the search span to find patterns related to the secondary retrieved patterns but may not be directly related to the initial search pattern.
The results panel provides an effective means to inspect the newly found behaviour classifications. Results are accepted by moving them to an appropriate classification tab in the classification widget, or rejected by clicking the reject button. The reject button removes the result from the panel. The user should aim to keep accepting/rejecting results until this panel is empty.
We appreciate that some instances will never be identified by machine learning and may only be able to be extracted by the domain expert, be that because of a low number of instances of the behaviour, or because of the variability of the animal behaviour. We enable manual labelling in our system so that the user can manually select and classify behaviour regions. To classify a behaviour region in the data manually, the user cuts the time-series graph up into segments. Each cut contains a start and end boundary defined by that of a behaviour instance. Once a behaviour region is cut in the time-series graph, the user drags and drops the time-series segment into the appropriate classification tab. The cutting and dragging of data samples is similar to that used in video-editing software.
Classified behaviours are shown in two views. Firstly, the classification widget which displays classified behaviour in a corresponding tabulated view. Secondly, classified instances are aligned and overlaid on top of time-series graphs as coloured rectangular regions identifying where in the data a match for the behaviour has occurred. Each behaviour is identified by a unique colour assigned to each classification tab in the classification widget.
Derivation of animal energy expenditure
Derivation of the physical characteristics of the environment
Many animals modulate their behaviour in the environment according to its physical characteristics. For instance, reptiles may associate with areas of high temperature or insolation to warm up  while many bird species are limited in their foraging capacities by light (e.g. ). Thus, the ability to resolve the geographic position of animals in tandem with environmental variables can help explain the incidence or emergence of particular behaviours (e.g. ). A key part of Framework4 is that it resolves animal space use and behaviour over time, inserting the positional and behavioural data into the original data file in columns. Given that DDs also record environmental variables, such as temperature , this means that these environmental variables are consequently linked temporally and spatially to the behaviour.
The derived analytical attributes from the software can be outputted together into one data file. Exporting data is supported via navigation to the ‘Export’ option in the ‘File’ menu on the main tool bar. Export is undertaken in CSV format where derived attributes are appended as an additional column in the data file alongside the existing data channels. Each behaviour is assigned a unique numerical value where, if a data item falls within a labelled region, it is assigned this value.
Results and discussion
The dead-reckoning and behaviour analysis components of Framework4 have been validated and compared with existing state of the art methods through two journal papers, in their respective fields. We refer the readers to Bidder et al.  for the verification of the dead-reckoning aspect of the software for animal movement tracks. The authors compare the dead-reckoning algorithm alongside the corrected dead-reckoning algorithm to geo-referenced data from video capture (using computer vision) along with time-aligned GPS, accelerometer, and magnetometer data. This is used to report several derived paths for comparison, dead-reckoned trajectories, GPS, dead-reckoning corrected using GPS, and dead-reckoning corrected using video at different sampling rates (refer to video in supplementary material). From this data, an error taxonomy model for the application of position correction is introduced.
Walker et al.  cover the behaviour labelling aspect of the software, presenting two case studies with domain experts applied to data obtained from turtles, and condors to determine specific behaviour events in the data. Using the software feedback loop, the experts were able to achieve 99.86 % and 100 % accuracy, respectively, in a short space of time. A direct comparison with machine learning techniques (hierarchical clustering, KNN, SVM, and random forests) is also presented using purely the algorithmic component of the software Walker et al. . Framework4 achieves a faster run time and accuracy in comparison to these methods. A visual analytics user-in-the-loop approach (as shown in the case studies) further lets accuracy be boosted to 100 % using domain feedback through the software’s user interface, something which was previously not possible. Framework4 allows biologists to achieve unprecedented levels of accuracy for their smart sensor data using a real-time interactive algorithm.
This is a first attempt to create a single program that explicitly links space use, movement, behaviour, and energy expenditure in free-living animals together with environmental conditions, doing so using an accessible column-separated format for ASCII-type data. Although there is appreciable room for improvement in many facets of the program at the moment, the aspiration is to progress and refine it to make it as powerful as possible and thereby provide a methodology which will enhance our understanding of the processes that affect the way that animals move within their environment. In future work, we seek to investigate segmenting movement tracks, applying vertical movement to dead-reckoning to generate three-dimensional movement traces, as well as investigate further visualisation techniques for movement data to expose patterns and trends in movement and behaviour.
Availability and requirements
Project name: Framework4
Project home page: http://www.framework4.co.uk
Operating system(s): Microsoft Windows 7 or newer. 64bit and 32bit windows supported.
Programming Language: C++ with Qt5.
License: The program was developed by JSW and is owned by Swansea University. We encourage its free use, no permission or license is required. The current paper should be cited in resulting publications.
Any restrictions to use by non-academics: None.
JSW was funded by an EPSRC doctoral training grant. IEM and EAM were funded by PhD studentships from the Department of Employment and Learning (DEL) and the Department of Agriculture and Rural Development (DARD), Northern Ireland, respectively. HW benefitted from a Swansea University PhD studentship. The Andean condor work was conducted in collaboration with Sergio Lambertucci, it was funded by National Geographic (grant GEFNE89-13 awarded to E Shepard), CONICET (grants PIP-0095, and ANPCyT–PICT-1156 awarded to S Lambertucci) and Swansea University, and data were collected with the help of a large field team.
- Aguilar Soto N, Johnson MP, Madsen PT, Díaz F, Domínguez I, Brito A, et al. Cheetahs of the deep sea: deep foraging sprints in short‐finned pilot whales off Tenerife (Canary Islands). J Anim Ecol. 2008;77(5):936–47.View ArticleGoogle Scholar
- Battaile (2014). “TrackReconstruction: Reconstruct animal tracks from magnetometer, accelerometer, depth and optional speed data”. http://cran.r-project.org/web/packages/TrackReconstruction/index.html
- Bidder OR, Walker J, Jones WM, Holton DM, Urge P, Scantlebury MD, et al. (this volume). “Step by step: reconstruction of terrestrial animal movement paths by dead-reckoning”. Submitted to movement ecology. Movement Ecology. 2015;3:23.Google Scholar
- Bidder OR, Soresina M, Shepard ELC, Halsey LG, Quintana F, Gómez-Laich A, et al. The need for speed: testing acceleration for estimating animal travel rates in terrestrial dead-reckoning systems. Zoology. 2012;115:58–64.View ArticleGoogle Scholar
- Bidder OR, Qasem LA, Wilson RP. On higher ground: How well can dynamic body acceleration determine speed in variable terrain?”. PLoS One. 2012;7, e50556.View ArticleGoogle Scholar
- Bidder OR, Campbell HA, Gómez-Laich A, Urgé P, Walker J, Cai Y, et al. Love Thy Neighbour: Automatic Animal Behavioural Classification of Acceleration Data Using the K-Nearest Neighbour Algorithm. PLoS One. 2014;9(2), e88609.View ArticleGoogle Scholar
- Börger L, Dalziel BD, Fryxell JM. Are there general mechanisms of animal home range behaviour? A review and prospects for future research. Ecol Lett. 2008;11:637–50.View ArticleGoogle Scholar
- Cai, J., Andersen, N. L., & Malureanu, C. (2001). “In-Field Practical Calibration of Three-Axis Magnetometers.” Proceedings of the 2010 International Technical Meeting of The Institute of Navigation (pp. 67-73).Google Scholar
- Caruso MJ. Applications of magnetic sensors for low cost compass systems. IEEE, San Diego, CA, USA, Honeywell, SSEC: Position Location and Navigation Symposium; 2000. p. 177–84.Google Scholar
- Chelazzi G, Calzolai R. Thermal benefits from familiarity with the environment in a reptile. Oecologia. 1986;68:557–8.View ArticleGoogle Scholar
- Ellis K, Kerr J, et al. “A random forest classifier for the prediction of energy expenditure and type of physical activity from wrist and hip accelerometers”. Physiol Meas. 2014;35(11):2191–2.View ArticleGoogle Scholar
- Esling, P and Agon, C. (2012). “Time-series data mining” ACM Computing Surveys. 45, 1, Article 12.Google Scholar
- Gao L, Campbell HA, Bidder OR, Hunter J. “Web-based semantic tagging and activity recognition system for species’ accelerometry data”. Ecol Inform. 2013;13:47–56.View ArticleGoogle Scholar
- Gleiss AC, Wilson RP, Shepard ELC. Making overall dynamic body acceleration work: on the theory of acceleration as a proxy for energy expenditure. Methods Ecology Evolution. 2011;2:23–33.View ArticleGoogle Scholar
- Halsey LG, Bost C-A, Handrich Y. A thorough and quantified method for classifying seabird diving behaviour. Polar Biology. 2007;30(8):991–1004.View ArticleGoogle Scholar
- Halsey LG, Shepard ELC, Quintana F, Gomez Laich A, Green JA, Wilson RP. The relationship between oxygen consumption and body acceleration in a range of species. Comp Biochem Physiol A Mol Integr Physiol. 2009;152(2):197–202.View ArticleGoogle Scholar
- Halsey LG, Shepard ELC, Wilson RP. Assessing the development and application of the accelerometry technique for estimating energy expenditure. Comparative Biochem Physiology A. 2011;158:305–14.View ArticleGoogle Scholar
- Hawkes LA, Balachandran S, Batbayar N, Butler PJ, Frappell PB, Milsom WK et al. The trans-Himalayan flights of bar-headed geese (Anser indicus). Proc Natl Acad Sci. 2011;108(23):9516–9.View ArticleGoogle Scholar
- Hooker SK, Baird RW, Al-Omari S, Gowans S, Whitehead H. Behavioural reactions of northern bottlenose whales (Hyperoodon ampullatus) to biopsy darting and tag attachment procedures. Fishery: Bulletin; 2001.Google Scholar
- Ioalè P, Dall'Antonia P, Dall'Antonia L, Benvenuti S. Flight paths of homing pigeons studied by means of a direction recorder”. Ethology Ecology Evolution. 1994;6(4):519–27.View ArticleGoogle Scholar
- Johnson MP, Tyack PL. A digital acoustic recording tag for measuring the response of wild marine mammals to sound”. IEEE J Oceanic Eng. 2003;28:3–12.View ArticleGoogle Scholar
- Jonsen ID, Myers RA, Flemming JM. Meta-analysis of animal movement using state-space models. Ecology. 2003;84(11):3055–63.View ArticleGoogle Scholar
- Martiskainen P, Järvinen M, Skön JP, Tiirikainen J, Kolehmainen M, Mononen J. Cow behaviour pattern recognition using a three-dimensional accelerometer and support vector machines. Applied Anim Behav Sci. 2009;119(1):32–8.View ArticleGoogle Scholar
- Miller PJO, Johnson MP, Tyack PL. "Sperm whale behaviour indicates the use of echolocation click buzzes ‘creaks’ in prey capture." Proceedings of the Royal Society of London. Series B: Biol Sci. 2004;271(1554):2239–47.Google Scholar
- Mitani Y, Sato K, Ito S, Cameron MF, Siniff DB, Naito Y. A method for reconstructing three-dimensional dive profiles of marine mammals using geomagnetic intensity data: results from two lactating Weddell seals. Polar Biology. 2003;26(5):311–7.Google Scholar
- Narazaki T, Shiomi K. Reconstruction of 3-D path (ThreeD_path). 2010. Available: http://bre.soc.i.kyotou.ac.jp/bls/index.php?3D_path.
- Nathan R, Spiegel O, Fortmann-Roe S, Harel R, Wikelski M, Getz WM. Using tri-axial acceleration data to identify behavioural modes of free-ranging animals: general concepts and tools illustrated for griffon vultures. J Exp Biol. 2012;215(6):986–96.View ArticleGoogle Scholar
- Qasem L, Cardew A, Wilson A, Griffiths I, Halsey LG, Shepard ELC, et al. Tri-axial dynamic acceleration as a proxy for animal energy expenditure : Should we be summing values of caluclating the vector?”. PLoS One. 2012;7, e31187.View ArticleGoogle Scholar
- Resheff YS, Rotics S, Harel R, Spiegel O, Nathan R. "AcceleRater: a web application for supervised learning of behavioral modes from acceleration measurements." Movement. Ecology. 2014;2(1):25.Google Scholar
- Renaudin V, Afzal MH, Lachapelle G. “Complete Triaxis Magnetometer Calibration in the Magnetic Domain”. J Sensors. 2010;2010:967245.View ArticleGoogle Scholar
- Ropert-Coudert Y, Wilson RP. Trends and perspectives in animal-attached remote sensing. Front Ecol Environ. 2005;3(8):437–44.View ArticleGoogle Scholar
- Rowley JJ, Alford RA. Techniques for tracking amphibians: the effects of tag attachment, and harmonic direction finding versus radio telemetry. Amphibia-Reptilia. 2007;28(3):367–76.View ArticleGoogle Scholar
- Sakamoto KQ, Sato K, Ishizuka M, Watanuki Y, Takahashi A, Daunt F. Can ethograms be automatically generated using body acceleration data from free-ranging birds? PLoS One. 2009;4(4), e5379.View ArticleGoogle Scholar
- Shepard ELC, Wilson RP, Quintana F, Laich AG, Liebsch N, Albareda DA, et al. Identification of animal movement patterns using tri-axial accelerometry. Endang Species Res. 2008;10.Google Scholar
- Shepard ELC, Wilson RP, Halsey LG, Quintana F, Laich AG, Gleiss AC, et al. Derivation of body motion via appropriate smoothing of acceleration data. Aquat Biol. 2009;4(3):235–41.Google Scholar
- Smith SW. The scientist and engineer’s guide to digital signal processing”. San Diego, CA, USA: California Technical Publishing; 1997.Google Scholar
- Vandenabeele SP, Shepard EL, Grogan A, Wilson RP. When three per cent may not be three per cent; device-equipped seabirds experience variable flight constraints. Mar Biol. 2012;159(1):1–14.View ArticleGoogle Scholar
- Vandenabeele SP, Wilson RP, Wikelski M. New tracking philosophy for birds. Front Ecol Environ. 2013;11(1):10–2.View ArticleGoogle Scholar
- Walker JS, Jones MW, Laramee RS, Bidder OR, Williams HJ, Scott R, et al. TimeClassifier - A Visual Analytic System for the Classification of Multi-Dimensional Time-Series Data”. Forthcoming: Visual Computer; 2015.Google Scholar
- Wanless S, Finney SK, Harris MP, McCafferty DJ. Effect of the diel light cycle on the diving behaviour of two bottom-feeding marine birds: the blue-eyed shag Phalacrocorax atriceps and the European shag P. aristotelis. Mar Ecol Prog Ser. 1999;188:219–24.View ArticleGoogle Scholar
- Watanabe S, Izawa M, Kato A, Ropert-Coudert A, Naito Y. A new technique for monitoring the detailed behaviour of terrestrial animals: a case study with the domestic cat. Applied Anim Behav Sci. 2005;94(1):117–31.View ArticleGoogle Scholar
- Wikelski M, Kays RW, Kasdin NJ, Thorup K, Smith JA, Swenson GW. Going wild: what a global small-animal tracking system could do for experimental biologists. J Exp Biol. 2007;210:181–6.View ArticleGoogle Scholar
- Wilson RP, Grémillet D, Syder J, Kierspel MAM, Garthe S, Weimerskirch H, et al. Remote-sensing systems and seabirds: their use, abuse and potential for measuring marine environmental variables. Mar Ecol Prog Ser. 2002;228:241–61.View ArticleGoogle Scholar
- Wilson RP, Culik BM, Bannasch R, Lage J. Monitoring Antarctic variables using penguins. Mar Ecol Prog Ser. 1993;106:199–202.View ArticleGoogle Scholar
- Wilson RP, Liebsch N, Davies IM, Quintana F, Weimerskirch H, Storch S, et al. All at sea with animal tracks; methodological and analytical solutions for the resolution of movement. Deep-Sea Research Part Ii-Topical Studies Oceanography. 2007;54(3-4):193–210.View ArticleGoogle Scholar
- Wilson RP, Pütz K, Peters G, Culik B, Scolaro JA, Charrassin JB, et al. Long-term attachment of transmitting and recording devices to penguins and other seabirds. Wildl Soc Bull. 1997;25(1):101–6.Google Scholar
- Wilson RP, Wilson MPT. Foraging behaviour in four sympatric cormorants. J Anim Ecol. 1988;943–955.Google Scholar
- Wilson RP, Shepard ELC, Liebsch N. Prying into the intimate details of animal lives; use of a daily diary on animals. Endangered Species Res. 2008;4:123–37.View ArticleGoogle Scholar
- Wilson RP, White CR, Quintana F, Halsey LG, Liebsch N, Martin GR, et al. Moving towards acceleration for estimates of activity-specific metabolic rate in free-living animals: the case of the cormorant. J Anim Ecology. 2006;75(5):1081–90.View ArticleGoogle Scholar
- Wilson RPB, Danfeld CR, Adelung D. “People in Antarctica—how much do Adélie Penguins Pygoscelis adeliae care?”. Polar Biology. 1991;11(6):363–70.View ArticleGoogle Scholar