Technical Comments

Response to Comment on “High-resolution global maps of 21st-century forest cover change”

See allHide authors and affiliations

Science  30 May 2014:
Vol. 344, Issue 6187, pp. 981
DOI: 10.1126/science.1248817


Tropek et al. critique the Hansen et al. global forest loss paper in terms of its utility and accuracy. Both criticisms suffer from a miscomprehension of the definition of forest employed as well as the requirements of product validation. Utility of the product is enhanced through its integration with forest type, carbon stock, protected area status, and other ancillary data.

The Comment by Tropek et al. (1) affords an opportunity to restate the definitions of forest cover and change used by Hansen et al. (2) and to illustrate the utility of our global product in a value-added application along the lines of those recommended in our paper. We recognize that many practitioners have specific forest definitions that do not conform to the one employed in our study. For example, many national forest agencies employ a land use criterion that is not tied to forest cover and its change. The cycle of planting and harvesting defines the forest land use, whereas quantifying harvest as a loss of forest cover (as in Hansen et al.) is often not part of forestry accounting. Similarly, labeling both the clearing of an oil palm estate and the clearing of an intact rainforest as forest cover loss may not conform to the research framework of tropical forest ecologists. Earth observation images capture the distribution of the biophysical features of Earth’s surface, something more commonly referred to as land cover. Our land cover theme of interest for this study was the presence or absence of trees at the Landsat pixel scale, where trees were defined as “all vegetation taller than 5m in height.” We produced a percentage tree cover per Landsat pixel layer for the year 2000 and used it as a reference in examining forest change. Forest loss was defined as a “stand-replacement disturbance,” meaning the removal or mortality of all tree cover in a Landsat pixel. Forest gain was defined as the inverse of loss or the establishment of tree cover from a nontreed state within a Landsat pixel. According to this definition, forest loss dynamics include drivers of change ranging from mechanical removals to fire to storm damage to stand diebacks due to disease. Land use, or the socioeconomic activity associated with a given parcel of land, was not a consideration in our mapping, nor was a finer thematic disaggregation of forests into natural versus managed or any other classification of forest by type.

The issues raised by Tropek et al. include (i) counting agroindustrial tree cover such as palm oil estates as forest, (ii) counting reestablished tree cover as forest gain, and (iii) mapping vegetation shorter than 5 m as forest. Issues (i) and (ii) are not errors according to our definition of trees. Oil palm, rubber, eucalyptus, and other managed stands qualify as tree cover if taller than 5 m. If a parcel was cleared before our study period, but regrew during it, then we can only map forest gain. Of the 24 examples listed in table 1 of Tropek et al., 19 are not errors when applying the definition of forest cover from Hansen et al. Although Tropek et al. question the conservation value of regrowing commercial forests, such considerations were not the point of our study or a part of our definition of forest and its loss or gain. Our study is entirely signal-driven, targeting the spectral reflectance signatures of trees from 2000 through 2012 at the global scale. Tropek et al.’s assertion that the results are misleading arises from their failure to recognize the definition explicitly stated of the feature being mapped—forest cover. If they and other users interpret the results based on the forest cover definition specified, the results cannot be misleading.

The third issue raised by Tropek et al. represents a clear case of classification error—mapping as tree cover vegetation that is shorter than 5 m. Searching for and finding examples of these types of errors is a simple task, but inferring from such examples that the product underestimates forest loss by “tens of percents,” as declared by Tropek et al., has no statistical validity. When assessing the accuracy of a map product, the validation data should be obtained from a probability sample to support statistically rigorous design-based inference (3). We have performed such an exercise at the climate domain scale in quantifying the accuracy and biases of our loss and gain map products with results reported in the supplementary materials of Hansen et al. (figures S5 and S6 and table S5).

Tropek et al. also conclude that the local relevance and utility of the approach is “seriously compromised.” An example application of a generic forest loss product that directly addresses the topical area that most interests Tropek et al.—conservation value forest—already exists. National-scale forest cover loss data for the Democratic Republic of Congo (DRC) were combined with forest type to assess primary, secondary, and woodland loss in both terra firma and wetland formations and combined with biomass data to quantify aboveground carbon loss (4). Results showed that one-third of DRC forest cover loss from 2000 to 2010 occurred within primary forests. However, higher gross aboveground carbon loss occurred within secondary forests than within primary forests. This is one example of many possible value-added applications of the forest cover loss data provided by Hansen et al. The creative use of these data through their integration with forest type, land use, carbon stock, protected area, and other data sets is appropriate and recommended.


View Abstract

Navigate This Article