Technical Comments

Comment on “Tropical forests are a net carbon source based on aboveground measurements of gain and loss”

See allHide authors and affiliations

Science  11 Jan 2019:
Vol. 363, Issue 6423, eaar3629
DOI: 10.1126/science.aar3629

Abstract

Baccini et al. (Reports, 13 October 2017, p. 230) report MODIS-derived pantropical forest carbon change, with spatial patterns of carbon loss that do not correspond to higher-resolution Landsat-derived tree cover loss. The assumption that map results are unbiased and free of commission and omission errors is not supported. The application of passive moderate-resolution optical data to monitor forest carbon change overstates our current capabilities.

Baccini et al. (1) report net tropical forest aboveground carbon stock change from Moderate Resolution Imaging Spectroradiometer (MODIS) data and purport to capture all forest carbon dynamics resulting from both natural and anthropogenic processes. We believe their method and results overstate current monitoring capabilities and may confuse the global community of practitioners working to establish robust and defensible forest carbon monitoring systems.

Our methods for mapping tree cover (2, 3) are very similar to those of Baccini et al. for mapping aboveground carbon; we use the same passive optical data (they use a simpler set of annual composite images that do not capture phenological variation) and the same data-mining tool (regression trees in a random forest implementation). The approaches differ principally in the dependent variable. Our research characterizing global tree cover uses MODIS and Landsat passive optical sensing systems. Land cover, including tree cover, is directly sensed in passive optical Earth observations because the multitemporal, multispectral reflectance variations directly relate to biophysical surface properties and vegetation types. Forest carbon is a more challenging variable to characterize, requiring knowledge about the volume and carbon density of vegetation—parameters that are not directly related to surface reflectance. Although the empirical model built by Baccini et al. suggests a relationship between passive optical imagery and forest carbon, it is likely the very same signal we use in mapping tree cover. Tree cover is largely discriminated using a spectral signature that is a combination of high reflectance in the near-infrared, associated with dense vegetation cover, and low reflectance in other bands due to light extinction from canopy shadowing and absorption from healthy vegetation cover. There is no physical basis for passive optical reflectance data being sensitive to changes in forest carbon parameters such as diameter at breast height or stem wood density.

Land change maps derived from satellite data consist of accurately mapped change along with inevitable omission and commission errors, whether the mapped variables are categorical or continuous. For data derived from MODIS, accurate mapping of land change is limited by its coarse spatial resolution (~500 m), as substantial areas of subpixel land change are not reliably quantified. For example, Hansen et al. (4) showed that area estimation of stand-replacement tree cover loss from Landsat samples at 30-m spatial resolution was twice that of a MODIS-derived tree cover loss map product. If MODIS data are forced to reproduce all of the area detected by Landsat, change maps quickly turn into noise, or more correctly, unconstrained omission and commission errors. The use of independent probability-based reference data to assess error rates and bias is standard practice and has been recommended by the IPCC (58).

Baccini et al. attempt to track changes in forest carbon stocks directly with propagated model uncertainties presented in place of accuracy assessments. Admittedly, an accuracy assessment of forest carbon change is difficult and would need to consist of a probability-based sample of time 1 and time 2 carbon stock measurements at the scale of observation. Regardless, in a modeling approach like that of Baccini et al., map biases and associated omission and commission errors are undetermined. This is critical, as the mapping challenge is greater for Baccini et al., given a dependent variable (aboveground carbon change) that has a much more tenuous relation to the MODIS-derived independent spectral reflectance variables. Baccini et al. state that their map is free from bias because they use a “single predictive model” applied to carbon stock gain and loss estimation. However, the distribution of map commission and omission errors is a function not only of model choice, but also of spatiotemporal variations of forest change, sensitivity to different land cover dynamics, data quality, and optical reflectance relationships with aboveground carbon.

We compared our tree cover loss data with Baccini et al.’s carbon loss estimates for the period 2003–2014 to assess map correspondence (Table 1 and Figs. 1 to 3). We found that 69%, 72%, and 43% of estimated carbon losses for Latin America, Africa, and Southeast Asia, respectively, are not colocated with Landsat-scale forest disturbances, defined as any MODIS grid cell with ≥10% Landsat-derived tree cover loss. The carbon losses reported by Baccini et al. must therefore be due to factors other than observable land cover and land use forest disturbances. Unfortunately, there is no evidence in their paper or in the literature to explain such large-scale biomass change dynamics for regions without tree cover loss. Other anomalous results include more than 25% of carbon loss in South America being located within intact forest landscapes with no observable human or natural disturbance (9, 10) and more than 30% of tropical Africa’s aboveground carbon loss found in parklands of 10 to 30% tree cover (2). On the basis of our experience with MODIS (3, 4, 11), much of the carbon loss presented by Baccini et al. is likely commission error.

Table 1 Disaggregation of Baccini et al. aboveground loss by percent tree cover, tree cover loss, forest intactness, and colocated tree cover loss and gain.
View this table:
Fig. 1 Tropical America.

Except for some large industrial-scale clearings, mapped carbon losses within the Amazon Basin are largely uncorrelated with Landsat-derived tree cover loss. Left: Tropical America tree cover loss from Hansen et al. (2). Right: Aboveground carbon (AGC) loss from Baccini et al. (1). Top row: Extent of study area of Baccini et al. for tropical America. Middle row: 2800 km × 1500 km subset centered on Mato Grosso state, Brazil. Bottom row: 530 km × 290 km subset of Amazonas state, Brazil.

Fig. 2 Tropical Africa.

The pattern of smallholder forest clearing along rural roads in the Congo Basin (12), which includes carbon loss due to the expansion of agriculture into primary forest, is wholly absent. Left: Tropical Africa tree cover loss from Hansen et al. (2). Right: Aboveground carbon (AGC) loss from Baccini et al. (1). Top row: Extent of study area of Baccini et al. for tropical Africa. Middle row: 2800 km × 1500 km subset centered on the Congo Basin. Bottom row: 530 km × 290 km subset of Tshuapa province, Democratic Republic of Congo.

Fig. 3 Tropical Asia.

The documented difference in logging within the primary forests of the Malaysia/Indonesia border (13) is wholly absent in the carbon loss map. Left: Tropical Asia tree cover loss from Hansen et al. (2) Right: Aboveground carbon (AGC) loss from Baccini et al. (1). Top row: Extent of study area of Baccini et al. for tropical Asia (east and west edges truncated). Middle row: 2800 km × 1500 km subset centered the islands of Sumatra and Borneo. Bottom row: 530 km × 290 km subset of central Borneo along the Malaysia/Indonesia border.

Moreover, Baccini et al. state that carbon change per MODIS cell is a net estimate, resulting in conservative aggregate gain and loss values. To assess this supposition, we examined MODIS grid cells with >10% loss and >10% gain from Landsat (2). Areas of both loss and gain accounted for 6%, 7%, and 21% of Landsat-defined mapped tree cover loss for tropical America, Africa, and Asia, respectively. We found no appreciable carbon loss in the data of Baccini et al. for these areas (bottom row of Table 1), supporting the idea that net change is low or absent in areas of colocated tree cover loss and gain. However, we posit that for other tree cover loss–dominated areas, carbon loss should be present. For tropical America, Africa, and Asia, Baccini et al. estimate no carbon loss for 68%, 76% and 72% of Landsat-derived tree cover loss–dominant cells; we identify these as errors of omission (Figs. 1 to 3) (12, 13).

Despite the likely presence of substantial spatial error and claims of capturing carbon losses from disturbances beyond deforestation, the carbon loss estimates of Baccini et al. are lower than those of Tyukavina et al. (14), which relied on robust and statistically accurate area estimation methods. Given the lower overall carbon loss estimate of Baccini et al.—an estimate not accounted for by colocated gain and loss, and consisting of considerable commission error—it follows that their estimates of carbon losses over landscapes experiencing land cover and land use change are underestimated. On the other hand, if we accept that the carbon losses as depicted are accurate, their results overturn our current understanding of where the global climate change science and policy communities should be focused in mitigating emissions from deforestation and forest degradation (i.e., not focused on tropical forest clearing). According to Baccini et al., deforestation is a minority source of emissions and is not colocated with forests experiencing degradation (i.e., sites spread diffusely throughout tropical forests, including forests heretofore thought to be intact) (Table 1 and Figs. 1 to 3).

We know the limitations of MODIS data in mapping forest cover, including signal saturation within high canopy cover and the impacts of spatial resolution, atmospheric contamination, and sensor degradation (15). Even if these effects could be accounted for, map characterizations inevitably contain error and bias. Baccini et al. provide per-pixel estimates of carbon stock changes from 1 to >1000 tonnes per hectare—a capability not previously demonstrated experimentally, much less at scale. We believe that the results of their study should be considered with an appropriate level of scientific skepticism, and that other approaches should be used until carbon stock change mapping can be more definitively demonstrated. We advocate the use of carbon stock strata—that is, areas of known mean carbon and associated uncertainty derived using probability-based forest inventory or other data. When such carbon stock data are combined with accurate estimates of forest/tree cover area change, forest carbon change can be quantified without bias and with known uncertainty (14).

References

View Abstract

Subjects

Navigate This Article