All 10,000 sample pixels were initially screened by two independent experts, who assigned forest loss (0, 50, or 100%) to each sample pixel. Pixels identified as 50 or 100% loss were attributed with loss year, predisturbance forest type, and forest disturbance driver (see the “Definitions” section). Experts also recorded their confidence (high/low) separately for the presence/absence of forest loss, forest loss year, predisturbance forest type, and forest disturbance driver. After the initial screening, sample pixels with disagreement between the two experts and pixels marked as “low confidence” for any interpretation category were additionally rechecked with the help of a third expert. Major sources of uncertainty during sample interpretation and the ways they were addressed by the interpreters are listed in table S4. Additional checks were performed using auxiliary data sources for the following pixels regardless of their initial confidence level:

(1) Primary and mature secondary HTF pixels with Landsat-modeled year 2000 tree cover <90% (14);

(2) Primary and mature secondary HTF pixels with Landsat-modeled year 2000 tree cover (14) >90% and year 2000 tree cover height (53) <15 m;

(3) Young secondary HTF pixels with Landsat-modeled year 2000 tree cover height (53) >20 m;

(4) Primary and mature secondary HTFs and primary woodlands and dry forest pixels in DRC outside primary forest mask (33);

(5) Young secondary HTFs and secondary woodland pixels in DRC within primary forest mask (33);

(6) Pixels with forest loss year 3 or more years different from the global forest loss map (14).

Sample pixels were iteratively rechecked by interpreters using auxiliary data until consensus on the final pixel labels and confidence levels was reached.

Because we used the best available information for our reference sample classification (visual interpretation of Landsat time series and available very high resolution data), it is not possible to formally assess the accuracy of our reference classification by comparing it to the “truth.” In a sense, our current sample classification is the closest practical approximation to this truth in the absence of historic annual (2000 to 2014) ground surveys or time series of very high resolution data. We therefore can only indirectly assess the possible errors of reference sample classification by analyzing certainty flags for each sampled pixel. A total of 497 sampled pixels (5% of the total sample size of 10,000; 274 “no loss” and 223 “loss”) were classified as low-confidence presence/absence of forest loss during sample interpretation (fig. S6). Sample pixels with low-confidence presence/absence of forest loss were spread throughout the region but were somewhat clustered in the cloudy coastal regions, particularly in the DRC province Bas Congo.

Confidence level was recorded separately for each forest loss category (loss year, predisturbance forest type, and loss driver), for both high- and low-confidence sample pixels identified as loss. Years with the highest percentage of forest loss area coming from low-confidence sample pixels (potential commission error) were 2007 (34%), 2003 (25%), and 2002 (25%); years with the lowest percentage were 2011 (9%), 2006 (10%), and 2010 (10%). Annual estimates for 2008 to 2014 had, on average, a smaller proportion of area coming from low-confidence sampled pixels compared with 2001 to 2007 (14% versus 20%), which may be related to a better availability of cloud-free Landsat data (fig. S5) and the very high resolution imagery from Google Earth in the later years. For the forest disturbance drivers, percentage of area coming from low-confidence sampled pixels was the highest in the two smallest classes: mining (55%) and natural disturbances (48%), followed by semipermanent small-scale clearing for agriculture (24%), logging (14%), large-scale clearing for agriculture (14%), fires (8%), rotational small-scale clearing for agriculture (5%), road construction (4%), and commercial and residential construction (2%). These differences could be related to the higher ambiguity of definitions for some of the classes. For example, construction classes are the least ambiguous, since they usually occur in the vicinity with already built-up areas and have a distinct postdisturbance spectral signature (concrete and dirt). Mining also has a distinct postdisturbance bare ground signature, but artisanal mining typical for the region is likely to be confused with natural disturbances (e.g., river meandering). Among the predisturbance forest categories, young secondary HTFs, secondary woodlands, and plantations had the highest potential commission error (22, 15, and 22%, respectively), due to their likely confusion with non-woody vegetation (young fallows, tall crops, and shrub). Primary and mature secondary HTFs and primary woodlands and dry forests had lower potential commission error rates (7 and 6%).

Note: The content above has been extracted from a research article, so it may not display correctly.

Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.

We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.