Frank van der Most, October 2021
At the IUCN’s World Conservation Congress I ran into two applications of machine learning (ML) to guard protected areas against direct detrimental human activity. Both stories are success stories but also, both applications are in their early phases. In this post, I will compare the two and ask a couple of common sense questions that may be of interest for the long-term viability of the approaches. In short, there is a lot of potential value in these ML applications, but I wonder how well they will do once the poachers and loggers start adapting.
ML against poaching and illegal logging
The two cases of ML application focus on fighting and preventing illegal logging and poaching respectively. The system against illegal logging is called an early warning system (EWS) to predict and prevent deforestation, and was presented by Jorn Dallinga, digital innovation lead at World Wide Fund for Nature (WWF) the Netherlands. The system against poaching is called Protection Assistant for Wildlife Security (PAWS) and was presented by Lily Xu, PhD student at Harvard University. Full references and links can be found at the end of this post. The information provided in this post is based on Xu’s and Dallinga’s presentations, which I attended and re-viewed for the writing of this post. A short introduction to machine learning is available in the post Machine Learning for dummies.
Big areas, few resources
The main reason for developing both systems has to do with effectiveness and efficiency. Protected areas are often big and understaffed. Xu provides the example of the Srepok Wildlife Sanctuary in Cambodia, which is 3724 square kilometers and has 72 rangers to protect and manage it. Because of the personnel shortage, it would be effective and efficient to be able to predict where illegal activities will take place and send out a patrol to those areas. Keep in mind that sending out a patrol is not a minor undertaking. They may take days or weeks, so it makes sense to give a lot of thought to the planning and the targets of a patrol.
Although the ML algorithms of the two systems are different, both approaches use historical data of the past years to base their future predictions on. Also, they both take into account a variety of geographical information sources: topographical data about where roads, rivers, villages are, the types of land coverage and more. The exact lists were not provided during the presentations, but it would be interesting to compare them.
Two differences stands out though. Firstly, PAWS is aimed to be integrated into SMART, a freely available, open source computer system for the integration of data from many different sources to support protected area management. EWS does not, or at least was not reported to be.
Secondly, the main source of forest degeneration and logging in the case of EWS are satellite based radar images. The main source for actual poaching in case of PAWS consists of patrol results and reports.
The resulting predictions of both systems consist of maps of the protected areas with an overlay of square tiles that indicate a probability of future (illegal) activity – possibly a bit like the image above. With these maps, area managers can develop their decisions for sending out patrols. The maps do not necessarily instruct the managers. Other considerations may come in as well. In case of EWS for example, not all logging that the machine predicts, necessarily is illegal logging.
At the time of presentation at the 2021 IUCN World Conservation Congress, both systems had been fully developed. That is the machines had been programmed, and their respective models trained and tested. Both have also been tested in the field: PAWS mainly in Cambodia, EWS in Indonesia, Gabon and Suriname, with five additional countries in the pipeline. Both systems have been very effective in their predictions and park rangers.
Given the presentations, the high success rate of the machines seem rather underwhelming. The creation of roads and canals in forest often precede logging and poaching usually occurs where most of the animals usually are. Duh! Does one need a complicated machine to analyze all this data to come to quite common-sense type of predictions? In the presenters’ defense, the machines are able to take in a lot more data (both in higher quantity and higher variety) than just the common sense indicators and are thus able to predict more precisely and perhaps also less predictable (i.e. less common sense) but still correct. The thing is that the common sense indicators are the only ones that stick out to humans and thus are easy to present to an audience.
Predict, patrol and prevent
Both systems are meant to predict illegal activity so that patrols can be send out to stop them, Dallinga pointed out that that is not the only action that can be taken. For example, when a certain area is highlighted as a high-risk area, but no illegal activity is actually going on (yet), measures could be put in place to prevent it from happening. Such as installing cameras or camera traps. Also, local people and local communities are often somehow related to the illegal activity. Instead of or in addition to policing and punishing, also discussions could be organized with local communities to try to prevent the illegal activities altogether. Problems with the boundaries of protected areas may be solved or alternative livelihoods developed, according to Dallinga.
What about the long term?
Both ML applications do not yet have a long life. I am very curious how things will develop over years of use. In particular, I am curious about the following. I would assume that the poachers and the loggers will at some point notice that the rangers start showing up a lot more often than before. They may not be aware of how and why, but they may still adapt to the situation. They may use different strategies and methods in order to prevent detection or become less predictable.
The next question is, if they do so, then how will the predictive machines respond? At this point a difference between the two presentations (although not necessarily between the two systems) stands out. Xu does not go into this topic, but Dallinga points out that EWS receives feedback about its predictions. This may cause it to change its predictive model and respond to the changes in the illegal loggers’ strategies and methods. The loggers may again respond and so on. A kind of cat-and-mouse game may develop. Perhaps I am overlooking something here, but in any case, I would be curious about the long term use of these predictive systems.
Efficient, effective and scalable?
The stated aims of the systems are to more efficiently and effectively deploy patrols to fight and prevent illegal activities in protected nature areas. The field tests indicate that these aims can indeed be achieved. The long term deployment of the systems may however show a decline in effectiveness if the cat-and-mouse game develops. On the other hand, long term deployment may also show that the illegal activities diminish or disappear from the protected area altogether. In that case, one should wonder where they go and does it mean a net benefit on a grander scale? Or does the investment in an ML system in one area simply means that illegal activities move to another, less well protected area?
This question (also know as the water bed effect) plays a role in all approaches to protecting nature. In case of ML systems, proponents may point to the scalability of computer systems. They are not all by nature scalable, but it is possible to design them that way. There is a cost to up scaling but the idea is that the costs grow remarkably less fast than the extension of the area under surveillance. That may be true for the computer systems, it may not be true for the costs of the patrols and other activities on the ground.
Patrols versus satellite imagery as a source of (feedback) data
I am not familiar with the price and availability of satellite radar images, but when it comes to efficient and effective tracking of illegal activities, they may in the long run outperform sending out patrols. To establish the effectiveness of the systems, and to enable the ML machines to continue learning, one needs feedback on their performance. False positive predictions can be discovered by new satellite images (in case of EWS) and by the patrols that are based on the prediction (in case of both PAWS and EWS). False negative predictions can be discovered by satellite images, but not (or less well) by the patrols that are sent out based on the positive predictions. Admittedly, these patrols also pass by locations with negative predictions, but this covers only a small percentage of the protected area. So, if the illegal poaching indeed moves to different areas as a result of the success of the PAWS system, that means that a number of patrols still need to be sent out at random, which in the long term could make the system less effective than advertised.
Does this make the satellite images a better source? It depends. In part, it depends on their acquisition costs, which I don’t know, and how they compare to sending out patrols. However, it also depends on the illegal activity that one wants to track and predict. Apparently, they can be used to find forest degradation, but perhaps they can not (yet?) be used to find animal traps, snares and people moving through the protected area.
Given the lack of resources to protect large areas of nature, it makes sense to look for efficient and effective ways of doing so. The introduction of technology to do so makes sense, but is not always easy because local customs, culture, work practices and even nature may need to be adapted. And the technology needs to be adapted to them. This also holds for ML learning systems. In fact, the point of introducing them is to change local behavior, i.e. to police and prevent poaching and illegal logging. The test results are promising, but that is not the whole story yet. The poachers and illegal loggers will adapt, which is also part of the introduction of the new technology. Hopefully they adapt by quitting their behavior, but they may respond in a different way. The future will tell.
Dallinga, Jorn (2021) ‘An early warning system to predict and prevent deforestation’. Session ‘The next generation of conservation data and technologies | How can innovative, digital technologies and citizen science help solve conservation challenges?’ at the IUCN World Conservation Congress. 6 September 2021. Session summary. YouTube (Dallinga’s presentation starts at 12:44)
Xu, Lily (2021) ‘PAWS: AI for Protected Area Management’. Session ‘Using Innovative Technologies to Improve Conservation Outcomes’ at the IUCN World Conservation Congress. 5 September 2021. Session summary. Replay (Xu’s presentation starts at 48:30).
Please note that to see the replay of Xu’s session, one needs to have registered for the conference.