Using Machine Learning to Combat Deforestation
Every year, tropical forest area the size of Austria is lost to illegal logging. The sheer size of forests and use of stealth tactics by loggers make it hard for local guardians to intercept illegal logging until it’s too late. Organizations like Rainforest Connection are using remote sensing technologies to monitor forests and detect illegal logging in real time. They work with forest communities to set up modified old cell phones fitted with solar panels in trees to monitor the forest. These acoustic monitoring systems process a continuous stream of audio data from the forest. Using extremely basic algorithms, they detect distinct sounds related to illegal activity such as chainsaws and/or logging trucks. When this happens, real time warnings are emitted to patrol members enabling them to intercept and stop illegal logging activities.
In application, acoustic monitoring devices have run up against limitations related to processing power and human capability. Consequently, Rainforest Connection is looking to machine learning to better use its data. First, each cellphone’s processing power limits its capacity to only basic algorithms. While these have proven effective in identifying distinct sounds such as a chainsaw, they cannot detect more sophisticated sounds such as human voices. This limits their monitoring capabilities. Second, the lack of programming capabilities restricts the ability of remote forest communities to improve monitoring algorithms and predictive models as increased data is received. Third, these constraints limit the potential applications of this large collection of forest audio data for other purposes, such as research. Given that Rainforest Connection is a non-profit, such other applications could provide sources for funding that would enable it to continue, and even expand its conservation work.
To address these issues, Rainforest Connection has partnered with Google’s TensorFlow open source machine learning platform. Processing data in the cloud solves the lack of human capability on the ground, since algorithms can be remotely adjusted. Moreover, the higher processing power of the TensorFlow framework can compensate for the lack of processing power in the devices, enabling detection of softer audio inputs (human voices) and improvement of detection quality. Furthermore, pattern recognition within TensorFlow can provide predictive insight into the behavior of illegal loggers, such as what areas they are most likely to target. Finally, TensorFlow is enabling wider applications of the data, such as using bird and animal sounds to detect the presence and movement patterns of endangered species like the Jaguar in the Amazon and improve conservation models.
The efficacy of the Rainforest Connection model can be further enhanced in two ways. First, by partnering with local universities and research centers, Rainforest Connection can add more regional capacity to work with forest communities. Moreover, enabling additional researchers to use data captured by the devices would increase the variety of machine learning tools and algorithms developed which could improve the overall predictive performance. Second, empowering the local forest community to use technology as a tool to facilitate conservation is essential. Rainforest Connection can focus on developing graphically focused interfaces to enable the community to use machine learning outputs more independently rather than permanently relying on Rainforest Connection programmers as a go-between.
As Rainforest Connection continues on this venture, two open questions remain. First, how can Rainforest Connection use machine learning to enhance the learnings from the data while keeping the basic model at a level that is accessible and understandable to the forest communities it serves? Second, is there a role for distributed innovation here, whereby regular citizens can work with the data in ways that might be helpful? Citizen science programs have been successful in developing species maps and gathering critical conservation data in the past (799 words).
 Centre for Global Development book, Why Forests? Why now? The Science, Economics and Politics of Tropical Forests and Climate Change (December 2016), by Frances Seymour and Jonah Busch
 White, Topher, “The fight against illegal deforestation with TensorFlow,” Google Blog (blog), 21 March 2018, https://www.blog.google/technology/ai/fight-against-illegal-deforestation-tensorflow/ accessed 12 November 2018
 Coldewey, Devin, “ Rainforest Connection enlists machine learning to listen for loggers and jaguars in the Amazon,” TechCrunch, 23 March 2018, https://techcrunch.com/2018/03/23/rainforest-connection-enlists-machine-learning-to-listen-for-loggers-and-jaguars-in-the-amazon/ accessed 12 November 12, 2018.
 Kelling, S., Lagoze, C., Wong, W., Yu, J., Damoulas, T., Gerbracht, J., . . . Gomes, C. (2013). eBird: A human / computer learning network to improve biodiversity conservation and research. AI Magazine, 34(1), 10-20. Retrieved from http://search.proquest.com.ezp-prod1.hul.harvard.edu/docview/1350239903?accountid=11311
 Balmford, A., and K. J. Gaston. 1999. Why biodiversity surveys are good value. Nature 398:204–205.
 Moran, Niklas ; Nieland, Simon ; Tintrup Gen. Suntrup, Gregor ; Kleinschmit, Birgit International Journal of Applied Earth Observations and Geoinformation, February 2017, Vol.54, pp.124-133