A Turning Point in Tropical Forest Monitoring
Brazil is a nation of superlatives. It is one of the most biodiverse countries in the world, and home to the largest rainforest, the largest freshwater preserve, and a quarter of the world’s species. It also one of the world’s largest grain producers and a major producer of coffee, sugar, and beef.
Brazil’s tropical forests are critically important for biodiversity, human livelihoods, and climate stability, but remain threatened. Despite zero-deforestation pledges from the government and agricultural companies, rapid deforestation in Brazil persists. Agricultural expansion continues to be the biggest driver of forest loss — driven by the world’s growing appetite for soy and other agricultural crops.
A New Frontier in Land Classification and Monitoring
Brazil was among the first countries to pioneer the use of satellite imagery to track annual land use change. However, these initiatives have been constrained by a lack of frequent, high-resolution data — making it difficult to segment different land use classes or detect degradation early enough to intervene.
This could soon change.
Recent improvements in the spatial and temporal resolution of satellite imagery by Planet and the availability of machine learning and imagery analysis at global scale by Google offer new opportunities to disrupt historical limitations.
The NexGenMap project is an R&D collaboration with Planet, Mapbiomas, Google, the Gordon and Betty Moore Foundation, to pioneer new practical and cost-effective solutions to track forest loss, classify change, and ultimately make more effective decisions about land use.
The project is focused on 16 areas of interest across seven distinct ecosystems in Brazil — a total of 300,000 square kilometers.
A Closer Look
One of the 16 sites is in Mato Grosso, an inland state in Brazil that historically has seen widespread forest clearing for agriculture, pasture, roads, and infrastructure.
The typical resolution of Landsat images is around 30 meters. While land use change is visible, in many cases it is not sufficient to understand the processes and factors driving deforestation in finer detail. For example, with Landsat imagery, it’s difficult to distinguish land use classes, such as pasture versus natural savannas, agriculture versus pasture, forest plantation versus natural forests. The low revisit rate of LandSat also makes early detection of degradation and deforestation challenging, often due to high cloud cover in the region.
Planet’s PlanetScope constellation images every location on Earth’s landmass every day at 3.7 meters resolution. As much of the tropical forests in Brazil are covered by clouds, this daily revisit rate enables Mapbiomas to generate broad-area mosaics without the worry of clouds and ultimately achieve more complete and recent analysis.
Real-Time Forest Monitoring
The mosaics using PlanetScope imagery are then processed by machine learning and deep learning algorithms developed by Mapbiomas and hosted in Google Earth Engine in order to automatically detect and map not only spatial patterns related to deforestation, but also segmentation of land use classes.
These automatic detections can be overlaid with third-party data, such as property boundary data (shown in white below) for more actionable tracking and enforcement.
Above is an example of automated deforestation alerts using Landsat imagery with property boundary data layered on top (see white lines). In this image from March 2018, the algorithms detected potential deforestation in properties A, B, C, and D (see red lines). The lower resolution makes it challenging for the algorithms to detect land cover change with precision.
In contrast, above is a PlanetScope mosaic from March 26, 2018 with property boundaries layered on top. Remember that in Landsat, the algorithms detected potential deforestation in A, B, C, and D. With higher resolution data from PlanetScope, you can confirm that deforestation occurred only in properties D and C.
With greater resolution, the algorithms can detect the deforestation more precisely, and with property boundary data, you can see where it is occurring and who is responsible. This can improve verification and enforcement of zero-deforestation commitments.
New Frontiers in Object Detection
Machine learning analysis of PlanetScope imagery also greatly enhances identification of common deforestation drivers, such as conversion to pasture and crops, and helps identify specific crop types. This is key, as understanding the drivers of forest loss is critical to better land management and policy.
Mapbiomas is currently exploring automated object detection of pivot irrigation for the purpose of differentiating natural grasslands and pasture lands. In addition, the project is conducting automated object detection of “bebedouros,” or cattle watering holes, that are often an early sign that land clearing for cattle will happen.
From Technology to Policy
Meeting human development needs, while protecting biodiversity and meeting climate goals, demands new practical solutions for monitoring land use change in Brazil and beyond.
Spatial and temporal improvements in satellite imagery, cloud computing, and machine learning are making it possible to detect and classify forest loss and land use change much more accurately and at scale.
Importantly, these technological advancements can be tied to monitoring systems that are used by governments, conservation groups, and decision makers to ensure better forest stewardship and enforcement. This will drive better outcomes for monitoring of commitments made by farmers, corporations, and financial institutions; land restoration investments; historical land use maps; and a new suite of land use tools to inform adaptive management and policy.
NexGenMap is graciously supported by the Gordon and Betty Moore Foundation.