Today, Global Forest Watch released a new forest monitoring system for Brazil called the GLAD alerts (named after the Global Land Analysis and Discovery lab at the University of Maryland). Brazil is certainly not new to the idea of monitoring forests via satellite, which may leave some users wondering about what this data shows and how it differs from other systems. Let’s delve into some of the technical methodology questions you may be asking yourself.
What do the alerts show?
As with most of the forest change data on Global Forest Watch, the GLAD alerts show areas of probable tree cover loss, not deforestation. While this difference may seem like semantics, not understanding this distinction can cause a lot of confusion. In the GLAD alert system, tree cover is defined as vegetation taller than 5 meters with a canopy density of 60 percent or greater. Each alert is a 30 meter by 30 meter area that has lost at least half of its tree cover. Deforestation, on the other hand, is the permanent conversion of natural forest into non-forest use, such as agriculture.
For example, harvesting in eucalyptus plantations is picked up as tree cover loss by the alert system, even though the change is temporary and not harmful to natural forests. Luckily, Global Forest Watch has a map of plantations in Brazil that can help identify which alerts show harvesting rather than clearing of natural forest.
GLAD alerts (in pink) pick up harvesting in eucalyptus plantations (in red). The tree plantations data can help identify which alerts happen within plantations rather than natural forest. View on GFW here.
In addition, not all of the tree cover loss is caused by humans—natural disturbances, such as wildfires, changes in river patterns and insect outbreaks, can cause loss that is picked up by the alert system.
GLAD alerts identify a potentially natural-caused fire in a remote area of Amazonas state. The burn scar in the UrtheCast image indicates that the loss is due to fire, but it is difficult to know whether or not there was human involvement without further information. View on GFW here.
Using other data, knowledge of the area or recent satellite imagery from UrtheCast can help distinguish human-caused deforestation from other types of tree cover loss.
How are the GLAD alerts produced?
Each week, the system compares newly available cloud-free and high quality Landsat images to images of the same area from the previous four years to identify any change. The algorithm detects changes by running the new images through seven decision trees for metrics, including ranks, means, and regressions of red, infrared and shortwave bands, and ranks of NDVI, NBR, and NDWI. The probability that each pixel is cleared is based on how much the result of this decision tree differs from previous years. Alerts are flagged when the probability of clearing is over 50 percent.
The GLAD alert system also determines whether or not each alert is “confirmed.” If the same alert is triggered in two different satellite images, it becomes confirmed. Unconfirmed alerts are discarded from the data after a second satellite pass since they are likely errors. Confirmed alerts have higher accuracy (more on that below), but since it takes two satellite images to confirm, which could take as little as 8 days or as long as several months to acquire, showing only confirmed alerts misses the newest detections.
You can read more about the methodology in the paper the team published in March.
How do the alerts compare to the Hansen annual data?
The GLAD alerts are produced by Matt Hansen and his lab at the University of Maryland, the same group that produces the annual tree cover loss data. The algorithms used to create the alerts and the annual data alerts are quite similar—the biggest difference is that the alerting system runs the algorithm weekly rather than yearly.
The definition of tree cover also differs between the two. While the annual data allows the user to select a tree canopy density to define tree cover, the alert only measures areas with 60 percent tree canopy density or higher since it was optimized to detect change in primary forests.
The alerts were designed primarily to serve as an early warning system indicating areas that have likely experienced recent loss, and underestimate the total amount of loss. The annual data provides a more definitive estimate of the amount of tree cover loss that has occurred within a year. Thus, the annual data is well suited for analyzing trends over time and the GLAD alerts can help support on-the-ground monitoring and enforcement efforts.
How accurate are the alerts?
So far, the only accuracy assessment of the alert system was done in Peru—there hasn’t yet been a published accuracy assessment for Brazil. In Peru, the alerts were found to have few false positives, but are generally a conservative estimate of tree cover loss.
The alerts had 13.5 percent false positives (loss detected where none occurred), though the majority of those false positives (9.5 percent) occur on the edges of clearings. On edges, the 30 meter Landsat pixels show a mix of forest and other land cover, which makes them prone to error in the system. The rate of false positives drops to 1 percent when only considering confirmed alerts. The data has 33 percent false negatives (undetected loss where it has occurred), though most of these occur in secondary forests—likely because the algorithm is calibrated to capture primary forest loss, which has slightly different physical characteristics. The higher rate of false negatives compared to false positives also indicates that the alerts are a conservative estimate of the tree cover loss that is actually occurring.
You can see all the details in the scientific paper.
Are they really weekly?
The GLAD alerts are updated weekly on Global Forest Watch, but the frequency of updates in any particular location depends on the availability of cloud-free Landsat images. Every location has a new Landsat image around every eight days, but cloud coverage may extend the time between observations to several weeks or even months.
Around 13 percent of Brazil has not had a cloud-free observation in the last month, and 5 percent of the country has not had a cloud-free observations in the last three months—cloud coverage will likely be even worse during the rainy season. Given this limitation, the absence of recent alerts doesn’t necessarily mean no new tree cover loss has happened, it may just mean that the images have been too cloudy for an update.
How do the alerts compare to DETER and SAD?
With so many different tree cover loss alerting systems in Brazil, it’s important to understand what sets each system apart. GLAD alerts are finer resolution than alerts by either DETER (created by Brazil’s National Institute for Space Research) or SAD (created by Brazilian environmental non-profit Imazon), cover all of Brazil rather than just the Amazon, and are updated weekly rather than monthly. However, the DETER and SAD systems use daily satellite imagery from MODIS rather than the 8 day cycle of Landsat, and so are more likely to get a cloud-free look than GLAD during the rainy season or in especially cloudy areas.
In addition, the accuracy of the GLAD alerts has not yet been determined in Brazil. It is possible that DETER and SAD are better at identifying loss in the Brazilian Amazon given that they were specifically made and calibrated for that area.
|Creator||University of Maryland Global Land Analysis and Discovery (GLAD) lab||Brazil’s National Institute for Space Research (INPE)||Imazon – Brazilian environmental non-profit|
|Coverage||All Brazilian biomes, plus other tropical countries (currently Peru, Republic of Congo, and Kalimantan)||Brazilian Amazon||Brazilian Amazon|
|Resolution||30 meters||250 meters||250 meters|
|Frequency of Updates||Weekly||Biweekly (internally), Monthly (externally)||Monthly|
GLAD alerts aren’t a perfect monitoring system, but they represent a new tool we can use to monitor forest change across the entire country of Brazil at a finer resolution and greater frequency than ever before. Combined with the ability to subscribe to your own areas of interest, monitoring change in Brazil just got easier than ever.