PREDICTING FUTURE POACHER MOVEMENTS BASED
ON PREVIOUS GPS LOCATIONS
by
Daisy Dwars
Student number: 12932434Thesis proposal submitted in partial fulfillment of the requirements for the Master of Science degree
in Earth Sciences
UNIVERSITEIT VAN AMSTERDAM WAGENINGEN UNIVERSITY & RESEARCH
Submitted to
Examiner: Mw. prof. dr. J.Z. (Judy) Shamoun-Baranes, professor at Universiteit van Amsterdam Co-assessor: Dr. ir. E.E. (Emiel) van Loon, associate professor at Universiteit van Amsterdam Supervisor: prof.dr.ir. F (Frank) van Langevelde, professor at Wageningen University & Research Supervisor: dr.ir. HJ (Henjo) de Knegt, associate professor at Wageningen University & Research
ABSTRACT
In most cases, anti-poaching rangers arrive too late at the crime scene. Being able to predict where a poacher would be in the foreseeable future, is thus extremely beneficial. More broadly, improving predictions is relevant in any field that involves the modelling of moving objects. The aim is to build a model for predicting the trajectory of a moving object in a heterogeneous area, where several environmental variables influence where the object will move to. A dataset from the SmartParks project will be used, which includes staged poacher trajectories. A machine learning method will be applied to build a model using R, with these trajectories as the input. The result of this thesis would be a machine learning model which is able to predict a moving object’s trajectory. In this case, a poacher’s trajectory for the next half an hour, with a one-minute interval.
1
1. INTRODUCTION
Being able to predict movement trajectories based on historical data of a multitude of different moving objects is of interest for a wide range of fields. For example, fishing vessels have an Automatic Identification Signals (AIS) system that records their location. AIS may drop out unintentionally, due to weak reception or interference. However, vessels with the intention of carrying out for instance illegal, unreported and unregulated (IUU) fishing may turn of the broadcasting intentionally to “go dark”. Predicting the location of such a vessel in the near future after it went dark allows the coast guard to intervene. Besides the fishing industry, making trajectory predictions is also relevant for the automotive industry as well as aviation. Making predictions can contribute significantly in creating reliable collision warning systems, enhancing the safety of autonomous vehicles, or preventing congestion (Wang, Wang, Chan and Fang, 2016; Ma, Zhu, Zhang, et al., 2019; Elfar, Talebpour and Mahmassani, 2018). In aviation, trajectory prediction is one of the most important duties in air traffic management (Shi, Xu and Pan, 2020). This can reduce flight delays as well as safety concerns for the flight of birds during take-off and landing of aircrafts for instance (Pang and Liu, 2020). Vehicles are not the only objects of which future locations can be predicted, human movement can also be modelled and predicted from historical data. For example, the location of an individual can be predicted based on previous visits to other locations (Mathew, Raposo and Martins, 2012). Anticipating human movement can be used for a whole range of undertakings. If human movement can be predicted, for example, then perhaps movement of criminals could also be predicted, for example poachers searching for target animals. With knowledge of prior locations of poachers, it may be possible to predict future trajectories of poachers.
Animals and plants traded illegally across borders, also known as wildlife trafficking, has become the fourth largest transnational organised criminal activity, after drug, arm and human trafficking (IUCN, 2019). Pangolin scales, elephant ivory and rhino horn are amongst the most trafficked products and numbers have increased to unprecedented levels during the last decade, with over 100,000 pangolins, 20,000 elephants and 1000 rhinos being poached each year (CITES report, 2019). Despite efforts to combat wildlife trafficking, it remains to be an incredibly lucrative business, with extremely low penalties when perpetrators do get caught (IUCN, 2019). O’Donoghue and Rutz (2016) outline how the fundamental problem is that in most cases, anti-poaching rangers arrive too late at the crime scene. Most anti-poaching strategies are treating the symptoms, namely seeking justice, yet knowing where the poacher would move to could aid in preventing the crime from happening. Accordingly, being able to predict where a poacher would be in the foreseeable future, results in extremely valuable knowledge. If it is possible to accurately predict future locations of a poacher, this means predictions could also be made for a range of other moving objects. This case study can generate new ideas which could then potentially be tested through similar methods. It is already known that it is possible to predict the movement of vessels and aircrafts, but it could go much further than that. The proposed model to predict poacher movements as a moving object will contribute to the literature by presenting an improved prediction model.
For predicting vessel and aircraft trajectories, course and speed are the main variables (Young, 2017). While cars are mainly limited to roads, vessels on open water and walking humans are less constraint to move around. Whereas vessels move in a more or less homogenous area, for predicting the trajectory of free-moving objects such as humans, a multitude of other environmental variables could play an important role, e.g. topography and vegetation cover are
2
crucial factors. Many studies have modelled trajectories of vehicles and the literature on predicting human movement is growing, however, not much is published on predicting poacher movement.
This thesis proposal therefore attempts to create a model for a moving object in a heterogeneous area where several environmental variables influence where the object will move to. Specifically, the model will predict the location of the poacher in the near future after they were first detected, as it takes time for anti-poaching rangers to arrive at the location. Accordingly, the research question is as follows: with what accuracy is it possible to predict the trajectory of a moving object in a heterogeneous area based on previous known locations? Additionally, how do environmental variables determine these future locations?
1.1 Theoretical framework
To address these research questions, a machine learning method will be used to build a model. For predicting future locations, machine learning methods can be split into parametric and non-parametric models (Shi, Xu and Pan, 2020). The former is based on prior known parameters, such as Markov Models, Time Series, and Autoregressive Integrated Moving Average (ARIMA) (Shi, Xu and Pan, 2020). These models generally require less data and are faster to train, however they make strong assumptions and are often unsuitable for more complex problems. Non-parametric models include Bayesian Networks, Neural Networks and Random Forests (Shi, Xu and Pan, 2020). These models are more flexible and make less assumptions, yet they require more data and are slower to train.
Young (2017) has attempted to predict the future location of a vessel from global AIS data. This was done by building two types of models in R, a Neural Network and a Random Forests model to perform regression. The latter proved to be more accurate. Zhang, Bin, Wang et al. (2020) also predicted vessel destinations based on AIS data, using a Random Forests-based model, where “the similarity between the vessel’s traveling and historical trajectories are measured and utilized to predict the destination”. Forti, Millefiori, Braca and Willett (2020) as well as Suo, Chen, Claramunt and Yang (2020) predict ship trajectories from AIS data using Recurrent Neural Networks. Suo, Chen, Claramunt and Yang (2020) find that a Recurrent Neural Network, performing classification, provides similar accuracy compared to a Long Short-Term Memory model, but it is computationally more efficient. For that reason, they state that the Recurrent Neural Network is more suitable for early warning systems.
Fernández, Cordero, Vouros, et al. (2017) use data from ground-based surveillance infrastructure or by onboard systems to predict flight trajectories, through performing regression. They defined and trained a Hidden Markov Model for four clusters in which the data was split, and found that the accuracy was highest for the larger clusters. Shi, Xu and Pan (2020) built a constrained Long Short-Term Memory network-based model for flight trajectory prediction using historical flight data. They compared this model to a number of other models and this method outperformed all the others. The use of sliding windows increased the accuracy of their prediction. Pang and Liu (2020) create a similar model; also using historical flight data they predict flight trajectories. However, they use the Bayesian Neural Network.
Elfar, Talebpour and Mahmassani (2018) predict short-term traffic congestion using historical data from vehicle trajectories using three different machine learning methods, Logistic Regression, Random Forests for classification and Neural Networks for classification. All
3
models had an overall accuracy ranging between 89% and 93% and they found that the accuracy was higher for a 10-second prediction horizon compared to a 20-second prediction (Elfar, Talebpour and Mahmassani, 2018). Ma, Zhu, Zhang, et al. (2019) propose a Long Short-Term Memory based real time traffic prediction algorithm performing classification, for autonomous cars to make predictions in relation to other traffic, allowing for safe navigation. Using historical traffic data, their model could predict with high accuracy trajectories of surrounding traffic. The accuracy varied however in different traffic conditions and the duration of past trajectories (Ma, Zhu, Zhang, et al., 2019).
Mathew, Raposo and Martins (2012) predict human mobility using a hybrid method of clustering and a Hidden Markov Model for classification. This model can “account with location characteristics as unobservable parameters, and also with the effects of each individual’s previous actions” (Mathew, Raposo and Martins, 2012). Gellert and Vintan (2006) predict human movement within an office building, using a Hidden Markov Model. With an average accuracy of 84.81%, they find that this model outperforms simple Markov predictors as well as Neural Networks.
What the previously mentioned models have not yet done, is the inclusion of environmental predictors in the model which in this case contribute to poacher movements. This thesis will add this extra layer of complexity. The model built by Young (2017) is most relatable to the proposed model for this thesis. Random Forests proved to be the most accurate machine learning method for predicting the future location of a vessel, hence this will be the principal method in this thesis. In case the output of the Random Forests model appears to be inaccurate, one of the above-mentioned models which have been used for predicting trajectories will be tried. Furthermore, while Random Forests is robust to outliers and scalable, individual trees are prone to overfitting. Other models have their drawbacks as well; Neural Networks require large amounts of data to train and Hidden Markov Models often have a large number of unstructured parameters.
4
2. METHODS
2.1 Data retrieval and processing
The data were collected and processed by the SmartParks Project team. By collecting GPS data from 135 sentinel animals (zebras and deer) with a tracking collar, staged poaching incidents could be detected with 86.4% accuracy based on the movement and behaviour of the sentinel animals (WUR, 2020). For this project, the staged poaching incidents were also tracked, i.e. the track followed by people emulating poachers. This resulted in a dataset of 116 trajectories, which all differ in length, yet they are on average about 45 minutes long (figure 1). The dataset includes time series data; the xy coordinates in UTM35S of trajectories of staged poachers with a one-minute interval. The staged poachers were either on foot, by car or hiding. There are 51 trajectories by car, 45 on foot and 20 where the staged poacher is hiding. It is not necessary to convert the coordinates as they are already in eastings and northings which are measured in meters rather than degrees. Some data processing will be done, for instance by creating separate data frames for the trajectories where the staged poacher was travelling by car, by foot or when hiding. It is worth mentioning that many of the above-mentioned studies include the haversine formula in their models, which accounts for the curvature of Earth. This will however not be included in this model, as the distances are small enough for it to be negligible. Geographic information system (GIS) raster layers with a 2-meter resolution in UTM35S projection will also be used. A mask of the study area, a digital elevation model (DEM), fractional grass cover, fractional tree cover and distance to nearest road (in meters and radians) layers are available in GeoTiff format.
5 2.2 Model
Time series data can be tricky to apply to machine learning, as most methods have no awareness of time. However, with some pre- and postprocessing of the data it is possible for e.g. Random Forests to make predictions. Using R, predictions for half an hour will be made with a one-minute interval. First the model will be trained and the hyperparameters will be chosen. By inserting the GPS data into the model, speed and course will be calculated in each observation, which will be used to predict the future locations. During the model training process, the model parameters will be learned and tuned. Afterwards, the model will be evaluated, evaluation metrics include prediction accuracy, generalization property, and complexity.
Increasing complexity
To start, a simple model will be constructed with only two variables. This model will then be elaborated for a more precise prediction to be made.
1. Constant speed and bearing
The bearing is the angle relative to true north that the poacher is traveling, from 0 to 359 degrees. The speed is how fast the poacher is moving, in this dataset either walking, by car or hiding. The bearing will be calculated using the following formula:
𝐴 = 𝑎𝑡𝑎𝑛2((𝑋1− 𝑋2), (𝑌1− 𝑌2)) (1)
where X1 and X2 are the eastings of a first and second location, and Y1 and Y2 are the northings.
A is the bearing in radians. Speed will be calculated using the following formula:
𝑆 = √(𝑋1− 𝑋2)2+ (𝑌1− 𝑌2)2 (2)
Where S is the speed in meters per minute, X1 and X2 are the eastings and Y1 and Y2 are the
northings of a first and second location. Speed will also be calculated for each observation. By taking the mean of speed and the circular mean of the bearing of each trajectory, while excluding n number of observations of the end of each trajectory, a prediction can be made and compared to the actual location. Making a prediction based on a constant speed and bearing will provide a prediction of a location which is most likely inaccurate. A next step can be to take a weighted mean of speed and bearing, which allows for the recent past to weigh more, as the start of the trajectory will likely have less influence than the last ten minutes on in which direction and with what speed the object will be moving to. However, building a model with solely bearing and speed will not provide accurate predictions.
2. Correlated random walk
As the poacher is probably not moving with a constant speed in one single direction, a second model could account for varying speed and course. Chances of the poacher continuing with the same speed in the same direction would be larger, while the chances are smaller for the poacher turning around 180 degrees (figure 3). In a random walk, an object takes steps at regular intervals where step lengths as well as direction are chosen randomly. In a correlated random
6
walk, a step forward is correlated with the direction of the previous step: the direction at step x+1 is related to the direction of step x. The turning angle is the change in direction for step x+1.
There are several packages that can be used for simulating correlated random walks. The SiMRiv package in R allows for specifying the amount of correlation the next step has with the previous step, by choosing a value between 0 and 1, using the state.CRW function (Porto and Quaglietta, 2019). Here, 1 means full correlation, leaving the direction unchanged, while 0 indicates a fully random walk is initiated where all directions have an equal probability for the next step. To incorporate the assumed tendency of a poacher to walk in a straight line, but not perpetually so, this parameter must be set to approximately 0.75. To account for variance in speed, a multistate movement can be defined. In the same package, the state.Resting function indicates a state where the object has stopped. Another package that could be used is adehabitatLT. In this package, using the NMs.randomCRW function, a correlated random walk can be simulated in which the turning angles can be randomized as well as “distances between successive relocations” (Calenge, 2020). Additionally, a constraint function can be added to prevent the simulation reaching areas outside of the park (Calenge, 2020).
It is also possible to create a correlated random walk from scratch, as shown in the R script in figure 2.
Figure 2: R script for a correlated random walk (Dwars, 2020).
An attempt will be made for the script of the correlated random walk to produce an output in the form of a heatmap as shown in figure 7. These probabilities of going into a certain direction can then be added into the Random Forests model as a third variable.
7
3. Naïve prediction
Young (2017) made a naïve prediction of the future longitude and latitude of the vessels. This is a “constant-velocity linear model that projects the current position in a straight line along the current heading using distance = rate x time” (Young, 2017). These rough predictions were then used as supplementary predictor variables in the Random Forests model, together with speed and course. It appeared that this variable was of high importance for predicting the future locations using the Random Forests model. This could also be done for the model proposed in this thesis. Furthermore, this linear model can be used as a null model. The null model can then be used to make a comparison to the Random Forests model.
4. Heterogenous environment
The environment has an influence on the speed and course. A fourth model would account for heterogenous environments. This means that more variables are included, namely topography, vegetation cover and infrastructure. For example, speed will be lower in areas with steeper topography or densely vegetated areas, whereas speed might increase when a poacher travelling by car finds better roads. The course is also influenced by the environment, as a poacher on foot might be more inclined to stay 20 meters parallel to a road, while poachers travelling by car might remain on the roads. Similarly, steep or densely vegetated areas might be avoided altogether. As white rhinos can often be found in grassy areas, poachers may be inclined to move to these areas (Pienaar, 1994). More research will be done as to where other target animals are usually found, which can then be taken into consideration. The aim is again to obtain a raster of probabilities portraying the likelihood of in which direction the poacher would move to, which can then be added as another variable in the model.
Random Forests
Random Forests is an ensemble learning method for classification and regression (Valletta, Torney. Kings, et al., 2017). In this model it will be used for regression rather than classification purposes. It is a data-mining algorithm for making predictions without overfitting. They are built from multiple uncorrelated decision trees, which consist of a root, nodes and leaves (figure 4). The root splits and forms branches, creating nodes. The final node is the node leaf. Random Forests uses bootstrap aggregation, also called bagging. To build each tree, a bootstrap dataset is made by randomly selecting samples from the original dataset, where the same sample may be used more than once (Magness, Huettmann and Morton, 2008). Predictor variables are the input variables, while the dependent variables are the variables being measured. Model predictions are averaged to obtain the final prediction, thus considering all possible outcomes. Samples that are not in the bootstrap dataset are called out-of-bag and are subsequently used to measure the prediction error. Therefore, cross validation is not necessary. For decreasing bias from correlation among trees, only a random subset of all predictors is considered while searching for the best predictor to use at each node (Magness, Huettmann and Morton, 2008). Furthermore, Random Forests provides an overview of the most important variables for making the prediction. In order of relevance for prediction, the variables can be ranked.
8
For the proposed model, the data will essentially be split into a training and validation set, where the training set is the bootstrap dataset, and the validation set is the out-of-bag dataset. About 80 percent of the data will be assigned to the training class. As mentioned above, n number of observations of the end of each trajectory will be excluded from the training dataset. The model will make a prediction of n minutes into the future and this output can then be compared to the actual location, i.e. the observations that were excluded from the model. For instance, the final ten minutes can be excluded, so that a prediction can be made for ten minutes into the future. For this model, it is assumed that course and speed will be among the most important predictor variables.
Figure 4: Random Forests regressor model (Nedjati-Gilani, Schneider, Hall, et al., 2017).
9
3. EXPECTED OUTPUT
The model will produce a set of predicted locations, for half an hour with a one-minute interval, with upper and lower bounds of eastings and northings. Figure 6 shows the true position as well as the upper and lower bounds of the latitude of each observation of every vessel two hours into the future (Young, 2017). Figure 7 shows a heatmap of the true and predicted position of a randomly sampled observation of the test set (Young, 2017). The predicted location is however not in the cell with the highest probability. The predicted position was derived from the validation prediction errors for the entire route. The model produced by this thesis will provide an output in a similar manner. Additionally, a table will be made outlining the model accuracies.
Figure 6: The true position of each observation of every vessel two hours into the future is represented by the black dots.
The upper and lower bounds of the latitude are represented by the red points (Young, 2017).
10
4. SCHEDULE
Table 1 provides an overview of the time schedule for this thesis. The aim is to finish writing the proposal in October. It will only be presented by December, due to earlier timeslots for the Research Workshop being full. Data preparation will take place from September to October, where after the programming will start in R to build a model. In this period, first a simple model will be made which will be made more complex by elaborating it step by step. By the end of January, the model should provide a satisfactory output. The writing of the thesis will already start in December, and the aim is to finish in February, so it can be presented by the end of February. This thesis is worth 36 ECTS, hence the project duration is scheduled to be six months. Although not included in the table below, weekly meetings are planned with the supervisors of the thesis.
Table 1: Proposed thesis time schedule.
September 2020 October 2020 November 2020 December 2020 January 2021 February 2021 Write proposal Presentation proposal Data processing Building model Write thesis Presentation thesis Duration
11
5. FUNDING
The costs for two supervisors, the examiner and co-assessor are in-kind, thus covered by the respective universities. The meetings will be held either at Wageningen University campus or online through Microsoft Teams, which is provided by Wageningen University. A Microsoft 365 Business Basic plan costs €4.20 a month per user (Microsoft, 2020). Nevertheless, this number is the upper limit and will be lower in reality as it can be assumed that Wageningen University has a deal with Microsoft. The train fare for a round trip from Amsterdam to Wageningen costs €37.5 (9292, 2020). Train fares are covered by the student travel product from Dienst Uitvoering Onderwijs (DUO). The data were collected by the SmartParks team are readily available. This means that there are no associated costs for obtaining the data. The data are stored locally for which no extra costs are required. For the preparation and processing of the data, R studio will be used which is a free software. Table 2 shows a breakdown of the necessary costs for this proposed thesis. The total expenses add up to €175.20.
Table 2: Proposed budget.
Month 1 Month 2 Month 3 Month 4 Month 5 Month 6 Total
Research costs (in €)
Supervision (UvA/WUR) In-kind In-kind In-kind In-kind In-kind In-kind N/A
Microsoft 365 Business
Basic (WUR) 4.20 4.20 4.20 4.20 4.20 4.20 25.20
Train fares (DUO) 37.5 0 0 0 37.5 75 150
Data 0 0 0 0 0 0 0
R studio 0 0 0 0 0 0 0
12
REFERENCES
Calenge, C. (2020). Analysis of Animal Movements: Package ‘adehabitatLT’. The Comprehensive R Archive Network.
Circular histogram of subject data. RPubs. Accessed 10 October 2020 at
https://rpubs.com/crmhaske/uwloo670_polarPlot.
CITES. Report on Monitoring the Illegal Killing of Elephants (MIKE). CoP18 Doc. 69.2 (2019).
Elfar, A., Talebpour, A., & Mahmassani, H. S. (2018). Machine learning approach to short-term traffic congestion prediction in a connected environment. Transportation Research Record, 2672(45), 185-195.
Gellert, A., & Vintan, L. (2006). Person movement prediction using hidden Markov models. Studies in Informatics and control, 15(1), 17.
Fernández, E. C., Cordero, J. M., Vouros, G., Pelekis, N., Kravaris, T., Georgiou, H., ... & Scarlatti, D. (2017). DART: a machine-learning approach to trajectory prediction and demand-capacity balancing. SESAR Innovation Days, Belgrade, 28-30.
Forti, N., Millefiori, L. M., Braca, P., & Willett, P. (2020). Prediction oof Vessel Trajectories From AIS Data Via Sequence-To-Sequence Recurrent Neural Networks. IEEE, 8936-8940.
IUCN. Wildlife Trafficking. Accessed July 2019 at https://www.iucn.org/regions/europe/our-work/wildlife-trafficking.
Ma, Y., Zhu, X., Zhang, S., Yang, R., Wang, W., & Manocha, D. (2019, July). Trafficpredict: Trajectory prediction for heterogeneous traffic-agents. Proceedings of the AAAI
Conference on Artificial Intelligence, 33, 6120-6127.
Magness, D. R., Huettmann, F., & Morton, J. M. (2008). Using random forests to provide predicted species distribution maps as a metric for ecological inventory & monitoring programs. In Applications of computational intelligence in biology. Springer, Berlin, Heidelberg.
Mathew, W., Raposo, R., & Martins, B. (2012). Predicting future locations with hidden Markov models. Proceedings of the 2012 ACM conference on ubiquitous computing, 911-918.
Microsoft Teams options. Microsoft. Accessed 29 November 2020 at
https://www.microsoft.com/nl-nl/microsoft-365/microsoft-teams/compare-microsoft-teams-options?market=nl.
Nedjati-Gilani, G. L., Schneider, T., Hall, M. G., Cawley, N., Hill, I., Ciccarelli, O., ... & Alexander, D. C. (2017). Machine learning based compartment models with permeability for white matter microstructure imaging. NeuroImage, 150, 119-135.
O'Donoghue, P., & Rutz, C. (2016). Real‐time anti‐poaching tags could help prevent imminent species extinctions. The Journal of Applied Ecology, 53(1), 5.
13
Pang, Y., & Liu, Y. (2020). Probabilistic Aircraft Trajectory Prediction Considering Weather Uncertainties Using Dropout As Bayesian Approximate Variational Inference. AIAA Scitech 2020 Forum.
Pienaar, D. J. (1994). Habitat preferences of the white rhino in the Kruger National Park. Proceedings of Symposium on Rhinos as Game Ranch Animals.
Porto and Quaglietta (2019). An R package for simulation and analysis of spatially-explicit individual multistate (animal) movements in any landscape: SiMRiv. The Comprehensive R Archive Network.
Reisadvies. 9292. Accessed 29 November 2020 at https://9292.nl/reisadvies.
Shi, Z., Xu, M., & Pan, Q. (2020). 4-D Flight Trajectory Prediction With Constrained LSTM Network. IEEE Transactions on Intelligent Transportation Systems.
SmartParks: using technology for anti-poaching. Wageningen University and Research. Accessed 29 November 2020 at https://www.wur.nl/en/project/SmartParks-using-technology-for-anti-poaching.htm.
Suo, Y., Chen, W., Claramunt, C., & Yang, S. (2020). A Ship Trajectory Prediction Framework Based on a Recurrent Neural Network. Sensors, 20(18), 5133.
Valletta, J. J., Torney, C., Kings, M., Thornton, A., & Madden, J. (2017). Applications of machine learning in animal behaviour studies. Animal Behaviour, 124, 203-220. Wang, P., Wang, J., Chan, C. Y., & Fang, S. E. (2016). Trajectory prediction for turning
vehicles at intersections by fusing vehicle dynamics and driver’s future input estimation. Transportation research record, 2602(1), 68-77.
Young, B. L. (2017). Predicting vessel trajectories from AIS data using R. Naval Postgraduate School Monterey United States.
Zhang, C., Bin, J., Wang, W., Peng, X., Wang, R., Halldearn, R., & Liu, Z. (2020). AIS data driven general vessel destination prediction: A random forest based approach.
