Southeast Asia is home to some of the largest, most diverse tropical forests in the world.
It is also home to some of the oldest forests, dating back an estimated 70 million years. Over the last few decades, many of these have been demolished at an alarming rate, with illegal logging and encroaching agricultural development proving difficult to surveil.
From the Harapan rainforest in Indonesia to the Banjaran Titiwangsaor Main Range in Malaysia, the rate and scope of deforestation was too great to be surveyed manually. Forest monitoring initiatives have begun to innovate, proposing new techniques and less labour-intensive methods of identifying deforestation over large swathes of land.
A 2018 study estimated that the region had lost 82,000 square kilometres of forest in the first 14 years of the century. Of this, another study published in Nature Climate Change, found that 12.8 percent of agricultural deforestation was due to plantations for pulp, forestry concessions accounted for 12.5 percent, just 2.1 percent was for mining concessions and 11 percent, in the years between 2000 and 2010, were due to industrial oil palm plantations.
A deeply rooted issue
Although palm oil is not the largest individual driver of regional deforestation, Global Forest Watch director Rachel Weisse said that its production yields shocking results locally.
“We see a lot of forests, primary forests, being cut down in places like Indonesia and Malaysia, in order to plant these plantations,” she said.
Indonesia and Malaysia together account for almost 85 percent of global palm oil production, with huge tracts of formerly thickly vegetated rainforest areas completely razed to support this output. But there is a strong economic dependence on the world’s most popular vegetable oil in these countries.
The product’s nearly ubiquitous presence in consumer products, from chocolate and bread to shampoo, cosmetics and candles means there is a seemingly insatiable demand for palm oil.
In 2009, the World Bank said that population and consumption increases meant the world would require an additional 28 million tonnes each year by 2020, on top of the 69.6 million tons produced in 2017/18.
Creating an automated forest guardian
Global Forest Watch, a project run by the World Resources Institute, has been monitoring the world’s forests using satellite imagery and web mapping tools since 2014.
In 2015, it received funding from the Generation Foundation, to develop a system for automatically identifying plantations in deforested areas using machine learning, the first prototype of its kind to use deep learning techniques on satellite imagery of the Southeast Asian forests.
According to Global Forest Watch, many palm oil companies in the region had zero deforestation commitments already, but Global Forest Watch manager Mikaela Weisse said they made it their role to provide data sets on these companies, making sure they kept up with their commitments.
“We’re seeing a lot more commitments being made by… oil palm companies that have
gotten a lot of public pressure,” Ms Weisse said.
“It’s really important for being able to act upon it and hold different actors accountable.”
Global Forest Watch said they used over 3,000 manually labelled images to train the algorithm, a process that was slow and tedious.
“[We] really just didn’t have the data and volume needed,” Ms Weisse said.
Collaborating with satellite imagery Orbital Insight, they fed their algorithm thousands of satellite images, teaching it to recognise the patterns, textures, colours and shapes of palm oil plantations. Orbital Insight provided the imagery and handled much of the labelling and storage of the imagery and working together, the slow process of labelling and supplying the images to feed into Orbital Insight’s algorithm was eventually completed.
Ms Weisse said the project utilised a supervised machine learning technique, hosted by a cloud-computing -based pipeline, which Orbital Insight had spent thousands of hours developing, in order to store the many petabytes of geospatial data. Global Forest Watch provided the team at Orbital Insight with a training guide on how to recognise a palm oil plantation. Thousands of individually marked satellite imagery samples from US-based imaging company Planet were fed to the convolutional neural network (CNN), teaching it to recognise areas that are plantations and areas that are not. These were extremely accurate, ground truth samples.
It trained the algorithm to recognise oil palm plantations by identifying patterns in colour, size and even texture. The tree lines, road networks and other identifying markers helped define the areas that the algorithm would come to automatically recognise.
This was an algorithm that saw patterns similarly to how a brain made connections, with very minimal pre-processing required.
Convolutional neural networks see patterns similarly to how a brain makes conceptual connections, with very minimal pre-processing of source material required.
They were designed to analyse visual materials, and to resemble the visual cortex of many animal brains. In other words, the CNN would learn the filters that would need to be classified manually in other algorithms, providing a major design advantage for Global Forest Watch’s application. By marking the imagery with areas that are oil palm plantations and areas that are not, the algorithm was able to recognise the characteristics automatically, recognising the shapes associated with a plantation, without a human having to input that information.
“Essentially it learns what those patterns are like, what kinds of factors are more indicative of an oil palm plantation versus not,” Ms Weisse said.
“The interesting thing about these convolutional neural networks is it’s a bit of a black box, so you know it works in very abstract ways… [but] don’t know what the factors actually are,” she said.
“Unfortunately the algorithm will not be easy to move to other sources of imagery,” she said.
“We would have to retrain it for each specific image-source.”
“Drone imagery also poses additional challenges given varying angles and heights of images captured”.
They acquired a vast amount of high-resolution satellite imagery from Planet, who provided almost daily image updates, which Ms Weisse said helped avoid training issues related to cloud cover.
“It’s resulted in being able to see almost all of our study areas in Southeast Asia without a problem, even though it is one of the cloudiest places in the world,” she said.
They would combine images from multiple days, ensuring their algorithm would always be able to analyse all the areas in the scope of their project, from Malaysia, Indonesia, Columbia and Cambodia.
Ms Weisse said that they wouldn’t be able to use other image sources for the project, relying on the high resolution images provided almost daily by Planet.
“We would have to retrain it for each specific image source,” she said.
“Drone imagery also poses additional challenges given varying angles and heights of images captured.”
Hurdles and new ways forward
The process was not free of challenges, and is not yet a perfect solution, however. The algorithm persistently confused different kinds of plantations, and the neural network had trouble consistently recognising plantations in the earlier years of their development.
Ms Weisse said that in Columbia in particular, they were picking up a lot of banana production areas in their results.
“We’ve talked a bit about if there are other sources of satellite imagery,” Ms Weisse said, but said that they hadn’t yet found the right solution.
In terms of the younger plantations, she said they would potentially look at more detailed training of the algorithm to recognise them at their younger stages.
Additionally, the deforestation processes that the model was detecting was already years into its development before the algorithm identified it, resulting in “a bit of a lag time”. This means that the tool is not yet responsive enough to stop deforestation as it occurs, but this could potentially be a future application — it still fulfils its use case of delivering supply chain transparency by being able to hold actors accountable.
Ms Weisse said there aren’t immediate solutions to these challenges, but that she hoped Global Forest Watch would be able to investigate them in future, adding that the team wasn’t new to overcoming technical challenges.
Ms Weisse said they knew using deep learning through a CNN would make monitoring the forests for palm oil plantations much easier, facilitating the identification of palm oil plantations with the bare minimum of human input.
Deep learning had been used in the past for all kinds of technologies, from facial recognition to language translation, but Ms Weisse said theirs was the first to use it to combat deforestation caused by palm oil production, a massive problem for Southeast Asia.
Ms Weisse said that the project was a landmark in terms of the use of technology to tackle one critical aspect of the multifaceted issue of excessive deforestation due to over development of palm oil, with plans to expand the project to Papua New Guinea, Liberia, Guatemala and Honduras.
“Our goal really is to use spatial data, especially from satellites to do better monitoring of the world’s forests,” she said.
