How do we know whether or not a REDD project is actually reducing deforestation and forest degradation? Satellite data is one increasingly popular answer. Computers can be trained to use the data to detect deforestation and changes in land use and plot the information on easy to read maps.
But there’s a catch. Astrid Bos is a PhD candidate at Wageningen University in the Netherlands, and works with CIFOR in Indonesia. Her work focusses on measuring carbon emissions from REDD projects. In a recent article on the CIFOR website, she explains that,
Different scientists train their computers in different ways. How sensitive should the computer be to changes in the vegetation? How should it distinguish between seasonality effects and “real” deforestation? Different decisions in the change detection process can therefore lead to different deforestation maps, even when they originally came from the same pair of ‘eyes’ – the Landsat 7 satellite.
The title of Bos’s article is “Can you trust the numbers?” The sub-heading reads, “Choice of data and methods greatly influences deforestation measurements for REDD+.”
Bos illustrates the problem with maps of deforestation from a mining area in Ketapang, West Kalimantan. The image on the left is from Google Maps showing the mining area. The image on the right is a close-up of two overlaid deforestation maps. The maps shows deforestation due to mining between 2001 and 2014. Both maps showed the areas coloured red as deforested. But the maps disagreed on whether the orange areas were deforested or not.
Bos is the lead author of a recent paper, published in the International Journal of Applied Earth Observation and Geoinformation. The other authors are Veronique De Sy (Wageningen University), Amy E. Duchelle (CIFOR Indonesia), Martin Herold (Wageningen University), Christopher Martius (CIFOR Indonesia and Germany), and Nandin-Erdene Tsendbazar (Wageningen University).
“Political manoeuvring around the data”
In their paper, the authors note that one approach to calculating carbon emissions from a REDD project is by multiplying the activity data over the project area by an emission factor. Activity data is the area of deforestation.
The authors write that,
The estimation of activity data evolved rapidly through innovations in remote sensing and forest monitoring, with algorithms and datasets with ever increasing levels of coverage, spatial and temporal detail, and accuracy. However, these datasets do not necessarily agree with each other, and more transparency and better cooperation between the science and policy domain is required to measure – and realize – the mitigation potential of REDD+ activities.
They list the possible problems: misalignment of reference levels and time periods; forest and deforestation definitions used; and data sources used for a map, such as different satellite data.
“The ambiguity,” they write, “leaves room for political manoeuvring around the data which threatens accountability.”
The paper compares two ways of measuring deforestation using satellite data: online map tools, that are easy to access and measure deforestation across the world; and open-source deforestation-detection algorithms, that are more customised to a particular location, but are also more time-consuming to use.
In their paper, the authors write that, “these datasets and tools are not always consistent or complementary, and their suitability for local REDD+ performance assessments remains unclear.”
The authors collected local reference data on deforestation in five REDD project sites: two in Indonesia (2 million hectares, and 3.6 million hectares); and one each in Peru (1.1 million hectares), Tanzania (200,000 hectares), and Vietnam (800,000 hectares). They then compared this data with maps from a global dataset, and a locally-customisable algorithm.
Inconsistent forest definitions
In Peru and Indonesia forest is defined as 30% tree cover. In Tanzania and Vietnam it is 10%. The minimum mapping unit also varies from country to country.
The authors note that in Vietnam’s 2016 REDD Forest Reference Level submission to the UNFCCC, converting forest to industrial tree plantations is considered as enhancement of carbon stock. For the purposes of their paper, conversion of forest to plantations is considered as deforestation because “at – at least – one point in time the forest was cleared which leads to a reflectance of bare soil”.
They found that neither dataset performed better overall. The locally-customised algorithm was more accurate in Tanzania and Vietnam. In Indonesia, the global map was more accurate. The difference between the two maps in Peru was negligible.
The authors write that,
The results show that both the magnitude and trend of deforestation delineated from the map products differed greatly from reference-based area estimates.
The authors conclude that one map might be better at measuring deforestation from mining, while the other is better at seeing changes from forest to oil palm plantations.
The authors experimented with different combinations of mapping strategies. They found that basing maps on the earliest detection of deforestation increased the accuracy.
But this only worked, “where both individual maps and tool were already reasonably good,” Bos writes. “One poor map can wreck the accuracy, even when combined with a good one.”
They recommend that, “More locally calibrated wall-to-wall products should be developed to make them more useful and applicable for REDD+ purposes.”
Can the deforestation data tell us whether REDD is working?
The authors compared deforestation over a period after the REDD projects started, with deforestation in the years before then.
They used estimates of the average annual deforestation before and after REDD at the five REDD project sites. These estimates were based on high-resolution reference data, but these estimates are subject to uncertainty, represented by confidence intervals.
If deforestation is reducing, that suggests that REDD is working. But as Bos points out, “if the estimated trend in deforestation is slight, and there is high uncertainty, it’s possible that in reality, deforestation levels could be stable or even increasing – raising questions about whether compensation is justified.”
In only one of the five REDD projects that they looked at, the authors found a “distinct downwards trend in deforestation regardless of uncertainty”. In the other four, “uncertainty in the trend remained”, Bos writes. They found that both datasets underestimated deforestation.
The authors compare Global Forest Change (GFC) data with the Breaks For Additive Seasonal and Trend (BFAST) algorithm. They write that,
In terms of REDD+ performance, these results reveal some ambiguity of the deforestation trends. In Peru, the GFC showed slightly increasing deforestation while according to BFAST deforestation was generally going down since the start of the REDD+ initiative. The site in Tanzania showed no clear performance while the steep drop in deforestation in site Indonesia-B after 2007 might indicate positive REDD+ performance.
The authors note that their objective was to look at the possibilities of combining datasets to increase accuracy, and to show the importance of uncertainties in the data. Their aim was not to calculate the change in deforestation to assess the performance of the specific REDD projects.
But the authors note that generally, “The direction and degree of REDD+ performance remained statistically uncertain as CIs (confidence intervals) were overlapping in most cases for the deforestation estimates before and after the start of the REDD+ interventions.”
Estimates of avoided deforestation should be ‘conservative’
Obviously, if REDD payments are to mean anything, they have to be based on reliable data. The authors note that if the data is uncertain, calculations of avoided deforestation should be conservative.
Bos raises the obvious question. “What does ‘conservative’ mean?”
At one of the REDD projects in Indonesia, the authors applied different degrees of conservativeness to the deforestation estimates. They found that varying ‘conservativeness’ and ‘confidence’ resulted in estimates of annual reduced deforestation between 7% and 20%. “A significant difference,” Bos comments.
Bos writes that,
These factors, plus the uncertainties in the maps, can therefore greatly influence the deforestation measurements reported under REDD+, and thus whether or how much compensation should be awarded.
Bos concludes that being “open and honest” about the strengths and weaknesses of the maps, tools, and reporting standards used should become common practice.
But the current reality is that the tools we have for measuring deforestation cannot be relied on to determine accurately the rates of deforestation in REDD project areas.