Mapping and monitoring carbon stocks with satellite observations: a comparison of methods

The monitoring requirements for reducing emissions from deforestation and forest degradation have been widely discussed and documented in a range of publications, including overviews of the general requirements to meet policy needs [1] as well as a variety of papers on the technical aspects and limitations of various monitoring approaches [2–5]. The general consensus of these documents is that monitoring of forest cover change using satellite remote sensing is practical and feasible for determining baseline deforestation rates against which future rates of change can be based, provided that adequate validation and accuracy assessments are conducted and documented. The type of monitoring and baseline approach used has been the subject of much discussion, with a range of modifications proposed to deal with equity issues among countries with different historical rates of deforestation. Methods to map and monitor forest degradation, in which only a portion of the forest stock is removed, have also been developed. These range from straightforward visual interpretation of satellite imagery at multiple spatial scales (grain sizes) [6] to semi-automated algorithmic techniques that require technical expertise to implement [7]. Mapping and monitoring of carbon stocks, on the other hand, has often been regarded as beyond the current capability of satellite remote sensing technology, despite great need [8], partly because much of the research on this topic has historically focused on field sampling approaches [9]. Nonetheless, mapping carbon stocks over large areas without satellite data is clearly problematic [10].

Basing UNFCCC (United Nations Framework Convention on Climate Change) REDD (Reduced Emissions from Deforestation and Degradation) policies on a carbon stock mapping approach would have a number of benefits relative to approaches based solely on field sampling and forest inventories. This is true not only in terms of improving estimates of carbon stored in forests for the emerging carbon markets, by providing spatially explicit information on the location of carbon stocks, but also with respect to avoiding the ambiguities, uncertainties and outright differences among land cover type classifications [11]. A carbon stock approach could allow countries to report at a higher IPCC (Intergovernmental Panel on Climate Change) reporting tier by providing country-specific data and advanced methods and data for land conversions, even at a Tier 3 level which is defined in the Good Practice Guidance (GPG) as including "models and inventory measurement systems tailored to address national circumstances, repeated over time, and driven by high-resolution activity data and disaggregated at sub-national to fine grid scales (GPG 3.17)" [12]. A carbon stock monitoring approach is directly linked with biomass dynamics. Since such an approach does not depend upon the determination of land use/cover types as a step in estimating biomass, the uncertainty associated with these classifications is removed. Land classifications may still be applied to the biomass map for the purposes of accounting and reporting, but they are no longer a necessary step in the determination of the biomass of the land area. A carbon stock monitoring approach will introduce different, and perhaps additional, sources of uncertainty than other more traditional methods, but can reduce the overall uncertainty level below Tier 1 methods, and depending on the specific situation, is also likely to reduce uncertainties below Tier 2 methods, in addition to providing geographically explicit information on changes in carbon stocks. This approach could be used to obtain estimates for above-ground biomass (AGB) in all of the categories of LULUCF (Land Use, Land Use Change and Forestry) reporting, including categories where land classification remains the same (i.e. Forest Land Remaining Forest Land) and in categories defining changes in land use (i.e. Land Converted to Forest Land). Additionally, this approach would also allow a Tier 2 key category analysis (GPG 5.30), as it can provide specific uncertainty estimates for each category measured with this approach.

In the remainder of this overview we use carbon stock and above-ground biomass terminology interchangeably (biomass is typically 50% carbon). We recognize that carbon stocks can refer to below-ground and soil carbon as well, neither of which are directly discussed here.

A number of approaches have been developed to map carbon stocks and AGB from the satellite observations described above. Each of the approaches relies on calibrating the satellite measurements to in situ estimates of AGB at field study plots. AGB is often determined using a combination of well documented allometric relationships between simple plot-level measurements (e.g. stem diameter, density and sometimes canopy height and/or depth) and AGB, where the latter is determined from trees that have been dissected, oven-dried and weighed [9, 24, 25]. This type of allometry has a long history and is used in daily forestry operations worldwide, although refinements are always needed and ongoing. We cannot describe all of the possible approaches to estimating AGB from satellite observations which include, for example, spatial interpolation and kriging of field measurements, but focus instead here on the most commonly used.

Comparisons and Potential for Monitoring

In addition to the general principles of the approaches, as described above, details of the data sets and specifics of the approaches used to produce the maps shown in Figure 1 are provided in the associated publications describing those maps [20, 27]. Both approaches made use of similar allometric equations [9, 24] (see 20 for more on this) relating the field measurements to estimates of AGB, but differ in the way those field data are associated with the GIS data layers (CA approach) or the satellite observations (DR approach). Comparing the maps visually shows the DR approach (Figure 1A) has more spatial detail across the region than the CA approach (Figure 1B), with both characterized by transitions between vegetation types ranging from dense humid forest in the Congo Basin to more open woodlands to the south. Note the more continuous nature of the DR map, and the more aggregated spatial units in the CA map. The comparisons to independent GLAS lidar metrics, which are closely related to field estimates of AGB [11, 17, 20] show a narrower range of variability within each AGB value in the DR approach, indicating less uncertainty at any given location in the DR map.

Even though a land cover map was not used (not needed) for the DR approach to estimate AGB, further comparison of the maps in Figure 1 was done by summarizing the AGB values in both maps by land cover type (Table 1). Although most of the classes show a similar average AGB values within each class of the Global Land Cover 2000 product [26], note how the averages based on the CA map tend to be higher than those derived using the DR approach (88% of the 25 cases). It is possible that the DR approach is systematically underestimating biomass within cover types, but validation of this map [20] and similar map products [29, 31–33] show that not to be the case. The apparent bias in the CA approach is most likely associated with difficulties in assigning field plot measurements to more generalized land cover categories, as described earlier. This is most evident in classes that are often modified by human land use (e.g. evergreen lowland forest classes), but also in areas with little or no vegetation cover (see last 5 classes of Table 1). Note also that the CA approach indicates higher AGB in deciduous woodland than in closed deciduous forest, and other apparent irregularities (e.g. between shrublands and grasslands).

Class Name	DR	CA	Δ
Closed evergreen lowland forest	216.3	273.5	57.2
Degraded evergreen lowland forest	121.2	171.5	50.3
Submontane forest (900 – 1500 m)	238.2	186.8	-51.4
Montane forest (>1500 m)	169.6	94.6	-75.0
Swamp forest	250.7	346.9	96.2
Mangrove	48.3	100.9	52.6
Mosaic Forest/Croplands	91.5	96.6	5.1
Mosaic Forest/Savanna	77.4	91.9	14.5
Closed deciduous forest	84.9	81.8	-3.1
Deciduous woodland	35.2	89.4	54.2
Deciduous shrubland with sparse trees	11.5	61.0	49.5
Open deciduous shrubland	12.8	61.6	48.8
Closed grassland	7.0	73.9	66.9
Open grassland with sparse shrubs	1.0	14.3	13.3
Open grassland	1.9	13.4	11.5
Sparse grassland	2.3	6.0	3.7
Swamp bushland and grassland	32.7	57.3	24.6
Croplands (>50%)	5.3	36.9	31.6
Croplands with woody shrubs	1.1	10.2	9.1
Irrigated croplands	1.6	44.8	43.2
Sandy desert and dunes	0.0	31.0	31.0
Stony desert	1.3	13.0	11.7
Bare rock	0.8	13.4	12.6
Salt hardpans	1.2	42.4	41.2
Water bodies	5.6	107.7	102.1

Land cover classes derived from the GLC2000, average AGB (tons ha-1) derived from the "direct remote sensing" (DR) and from the "combine and assign" (CA) approaches.

The column labelled Δ indicates the difference between approaches (CA-DR).

Any of the approaches described herein could potentially be used in a framework for monitoring AGB stock changes. An SM approach would use new land cover maps to estimate changes in stocks, but the maps would need to be recreated at both (each) time step at accuracies that exceed those of typical current map products (particularly given that errors multiply whenever categorical maps are differenced for two time periods). A CA approach could make use of successive field surveys to update maps derived using GIS models, but field data are unlikely to ever be sampled adequately for monitoring purposes, particularly outside "intact" forests [35], and the issues noted above with respect to use of progressive land cover maps (and other spatial data sets) also apply. Moreover, both of these approaches are prone to missing changes at grain sizes finer than the thematic map units, including forest degradation and afforestation (regrowth) at time scales that are meaningful to the UNFCCC process. These issues are unlikely to impact the DR approach for monitoring purposes because land cover maps are not required, the spatial unit is typically much smaller than the other approaches (i.e. the pixel size), and the data sets used (satellite observations) are not only more current but also more consistent through time (via routine calibration efforts). Unlike land cover or land use maps, the satellite observations used in the DR approach are sensitive to changes at the pixel level, thus even fine-scale changes associated with degradation and afforestation can be detected [11]. Nonetheless, more work needs to be done with the DR approach in order to refine the process, reduce uncertainty (error), and better quantify the requirements (in terms of sample sizes and spatial extents) for field data sets needed to calibrate and validate the models. This topic is the focus of both applied research for the UNFCCC process and basic research designing the next generation of satellite missions that will provide the essential data sets needed for any monitoring system relevant to reducing emissions from deforestation and forest degradation.

Well documented techniques and satellite data enable reliable mapping of carbon stocks over large areas. Although a number of relatively simple methods exist for assigning field estimates of AGB to categories defined by vegetation type classes or weighted data layers in geographic information system models, the most spatially consistent maps are produced using models that derive continuous values (e.g. between 0 and 500+ tons per hectare) from statistically optimized decision rules. The techniques and data sources described herein are published in the refereed scientific literature and are progressing rapidly as new "data mining" techniques are advanced [28] and improved satellite remote sensing data become available [11]. This situation will improve further as new satellite missions come online in the next few years, several of which are designed specifically with the intent of improving estimates of the standing stock of carbon in biomass, and changes in those stocks through time. The UNFCCC process would benefit from refinement and application of these approaches and from improved data in developing policies designed to reduce emissions from deforestation and forest degradation.