Methodologies Used by the Global Forest Cover Change Project
1. Fine Resolution Surface Reflectance ESDR
The fine resolution surface reflectance (SR) ESDR products will consist of orthorectified Landsat images converted to surface reflectance using radiometric normalization techniques including atmospheric correction for TM and ETM+ images and radiometric rectification for MSS images. Each image product will contain pixel level masks flagging cloud and cloud shadow. These products will be made using the four Landsat data sets included in the Global Land Survey (GLS) collection, with global coverage for 1975, 1990, 2000, and 2005. The SR ESDR will be produced through the following steps: orthorectification, calibration and TOA reflectance, atmospheric correction, and cloud and cloud shadow masking.
2. Fine Resolution Forest Cover Change ESDR
The fine resolution FCC ESDR suite will include a 57 m FCC record between 1975 and 1990 and 28.5 m FCC records between 1990, 2000, and 2005. We recognize that different users will have different minimum mapping unit requirements and hence we will leave it to users to aggregate these data sets according to their own individual needs. These products will be produced for all land areas of the globe, to the extent allowed by available fine resolution satellite data sets. We estimate that it should be possible to provide these products for approximately 80% of the forested areas of the globe between 1975 and 1990, 85% between 1990 and 2000, and near 100% between 2000 and 2005. For reasons discussed in section 3.1, only net forest cover type conversion changes will be included in the proposed fine resolution FCC ESDR products. Conversion from forest to non-forest is defined as forest gain and the reverse process as forest loss. Forested and non-forest areas that did not experience forest cover conversion changes during the concerned period are defined as persisting forest and persisting non-forest respectively. Based on the global land cover classification system developed by the International Geosphere Biosphere Programme (IGBP) (Belward & Loveland, 1996), this will include all areas having at least 30% or more tree cover.
A wide range of techniques exist for land cover change analysis using satellite data. Comprehensive reviews of those techniques have been provided in several publications (e.g. Singh, 1989; Lunetta & Elvidge, 1998; Coppin et al., 2004; Lu et al., 2004), (e.g. Gordon, 1980; Green et al., 1994; Lunetta et al., 2004) Many techniques have been tested with varying levels of success over small areas (e.g. Jha & Unni, 1994; Cohen et al., 1998; Lyon et al., 1998; Tokola et al., 1999; Franklin et al., 2002). However, only a few were used to produce FCC products over relatively large areas, and these required intensive human inputs in order to achieve satisfactory results (e.g. Skole & Tucker, 1993; Townshend et al., 1995; Loveland et al., 2002; Huang et al., 2007a). Producing globally consistent and reliable FCC products at affordable costs requires highly automated change detection algorithms requiring minimal human inputs. To meet this challenge we have developed a method integrating training data automation (TDA) and support vector machines (SVM).
Figure. Use of MODIS data (b) to adjust images acquired during different phenology stages (a) to a common leaf-on season (c) using the Gao et al. algorithm (2006). When used to develop FCC map, the original leaf-off image did not allow detection forest loss (d). But after being adjusted to a leaf-on season based on the MODIS data, the forest loss class was mapped successfully (e), and the change map was quite similar to a change map derived using an actual leaf-on image. The circled changes in (e) are real changes occurred after the acquisition of the actual leaf-on image and before that of the leaf-off image. In (d)-(f), persisting forest, persisting non-forest, forest loss, and forest gain are shown in green, light brown, red and cyan, respectivel
Previously we have used decision tree algorithms to develop land cover classifications at the global scales (Hansen et al., 1996; DeFries et al., 1998; Hansen et al., 2000). However, we are convinced by the theoretical advantage of SVM and its demonstrated superior performances that it will provide the best solution for global FCC analysis. Since training the SVM is now fully automated it is possible to automate the entire multi-temporal SVM classification approach to forest cover change analysis. To ensure and improve the quality of the initial FCC ESDR products, we will implement a consistency and quality checking process (CQC). This process is designed to highlight and correct potential errors in the FCC products.
3. Fragmentation ESDRs
While the developed FCC products can be directly used in climate models and biogeochemical models, many ecological applications are also interested in the impact of FCC on habitat, which is often characterized using fragmentation indices (also called metrics in the biodiversity community). Measures of fragmentation address impacts of areal and spatial alterations of land surface on aspects of ecological systems (Lovejoy et al., 1986; Bierregaard et al., 1992; Laurance et al., 1997). Of principal concern is how fragmentation affects landscapes as well as species and therefore the issues of migration, habitat and services.
A collection of fragmentation products will be derived from the fine resolution FCC ESDRs through calculation of standard fragmentation indices, including patch area (Patton, 1975), perimeter/area (Hargis et al., 1998), shape (Patton, 1975), core area (Temple & Cary, 1988), and patch cohesion (Schumaker, 1996). These indices will be generated to facilitate the preferred methods, and their combinations therein, within the conservation community (Bogaert, 2003). These indices characterize fragmentation from simple relationships of forest cover patches to overall area with "patch area" (Patton, 1975) through the complex measure of connectivity between forest cover patches in "patch cohesion" (Schumaker, 1996).
The fragmentation indices will be calculated for the three epochs where TM and ETM data are available. Scale-dependent measures will not be generated with the MSS products because of their different resolutions.
4 Moderate Resolution ESDRs
The VCF percent tree cover products are made using a regression tree algorithm to generate annual percent tree cover layers. Regression trees have been used in remote sensing by others (e.g. DeFries et al., 1997; Mickelson et al., 1998; Hansen et al., 2002a; Hansen et al., 2002b; Huang & Townshend, 2003). They are a non-linear tool which recursively splits a continuous training variable into subsets, called nodes, which minimize the overall residual sum of squares.
This approach was prototyped using AVHRR data sets (Hansen et al., 2002a; Hansen & DeFries, 2004), and is currently being implemented as a standard MODIS land product (MOD44b) (Hansen et al., 2002b). Currently 6 years of 500 meter spatial resolution VCF tree cover data are available (Pittman et al., in preparation) as deliverable ESDR products. A further refinement of the VCF tree cover maps is currently being implemented as part of NASA's Land Cover and Land Use Change (LCLUC) program in a project entitled "Establishing a global forest monitoring capability using multi-resolution and multi-temporal remotely sensed data sets." The project is employing the VCF annual layers and MODIS time-series data sets to identify global change hotspots in developing a high-resolution block sampling protocol for estimating change. The MODIS maps take the form of change likelihood layers, not direct area estimates, as sub-pixel change training data are not available at the global scale. This proposal will significantly improve on area change mapping with MODIS data by using the fine resolution change maps (section 126.96.36.199) directly as training data. Given a representative change training data set, the standard VCF methodology can be applied to map annual forest loss. As the fine resolution 2000 to 2005 change maps are generated, they will be incorporated into the VCF methodology to produce a global VCF tree cover change product on an annual basis.
5. Aggregated ESDRs
To facilitate the use of the developed ESDR products in carbon and hydrological models, we will aggregate them to coarse spatial resolutions. Since MODIS land products are tailored for use by modeling communities (Justice et al., 2002), we will use the grid size of those products, including 250 m , 500 m, 1 km, and 0.05°, to generate the aggregated products.
6. Protected Area ESDRs
Because of the importance of protected areas a subset of the FCC ESDRs for all officially recognized protected areas will be generated using the World Conservation Monitoring Centre's World Database on Protected Areas (WDPA). The WDPA polygon for each protected area with more than 10% forest cover will be used to subset the fine and resolution ESDR products. The subset will also include buffer zones of 10km following the work of Bruner et al. (2001) and 50km following Defries et al (2005)
Contact the GLCF for any questions regarding
the GFCC project.