A1. Approach for Mapping LUC
 The Land Use Change Mapping System for Canada's North (LUCMAP-N) is based on combining several normalization and land cover change detection techniques recently developed and tested by our team [Fraser et al., 2005; Latifovic and Pouliot, 2005; Olthof et al., 2005a, 2005b]. These techniques were designed to automate the change detection process as much as possible, while still employing expert analyst guidance and quality control. LUCMAP-N uses a hybrid change detection method, which combines two separate techniques: change vector analysis (CVA) for identifying changed areas and constrained signature extension for labeling of those changes. CVA is a radiometric change detection approach that identifies the magnitude and direction of change in multispectral space [Johnson and Kasischke, 1998]. Unlike many radiometric change methods, CVA is able to exploit the full feature space of the imagery and also incorporate derived channels, such as texture.
 One difficulty in using remote sensing for monitoring LUC is that the technology is better suited for providing information on the biophysical state of the land surface, such as land cover type [Loveland and DeFries, 2004]. Land use indicates the manner in which biophysical assets are used by people. While land use is often constrained by land cover, it is also determined by other factors that are economic, institutional, cultural, and legal [Cihlar and Jansen, 2001]. In some cases, the relationship between land cover and land use can be direct, such as built-up cover type corresponding to urban land use. However, more commonly a given land use can correspond to more than one land cover type and vice versa [Cihlar and Jansen, 2001].
 To make the land cover and land cover change products a better surrogate for land use, we exploited other, nonremote sensing geographic data sets to provide ancillary information. These data sets were visually interpreted along with the change detection products and contextual information (shape, size, autocorrelation) contained in the images to infer land use change. A flowchart of the LUC mapping steps is shown on the left-hand side of Figure 2, and a summary of each follows.
 Step 1 is preprocessing and image normalization. First, we warped 1985 and 1990 scenes to 2000 scenes. The PCI GCPWorks package was used to collect 15–25 ground control points to warp channels 3–5 of each scene into six empty channels contained in the LCC georeferenced 2000 scene. In all cases RMS error was between 0.2 and 0.5 pixels. To ensure that the scenes were radiometrically consistent, Landsat channels 3, 4, and 5 of the 1985 and 1990 imagery were normalized to the baseline 2000 imagery using Theil-Sen robust regression [Fernandes and Leblanc, 2005]. Theil-Sen is ideal for normalizing imagery because it is insensitive to up to 29% of outliers, which in this case could represent changed pixels. In addition, Theil-Sen provides a fully automated method of normalizing imagery, unlike conventional methods that require the analyst to interactively select invariant targets.
 Step 2 is to change detection. We applied CVA to Landsat channels 3, 4, and 5. A simple model was executed by an EASI change detection script to compute the CVA channel.
 Step 3 is to derive change threshold. The CVA magnitude was displayed onscreen to allow the analyst to interactively select an appropriate threshold to create a binary change/no-change mask. Since the resulting change area is highly sensitive to the chosen threshold, a liberal threshold was used that may have included some nonchange areas in addition to all changed areas. Subsequent steps were designed to identify and remove the unlikely changes included in this mask.
 Steps 4 and 5 are to apply signature extension for postclassification comparison. The rich interpretative information embedded in the northern 2000 baseline land cover served as the basis for labeling pixels identified as having changed state. This was accomplished extending the labeled land cover clusters (i.e., signatures) from a circa 2000 land cover product to the other dates using a minimum distance supervised classification.
 Step 6 is to convert land cover change to LUC. The final, and most labor intensive step was to convert the land cover change products and images into one that includes only the subset of changes representing LUC. These LUC areas were identified manually across the image, and heads-up digitized in ArcMap by creating shape files for polygon and linear features. While editing the LUC shape files, a decision to digitize a land cover change as a LUC was made based on several lines of evidence: the original RGB TIFF images for each date to examine the spectral and contextual patterns in changed areas; RGB difference images that show the spectral trajectory of change; CVA magnitude and change mask channels; the land cover classification for each date; a large database of recent digital aerial photos from the Mackenzie Air Photo Project; the relation of the land cover changes with respect to our database of the ancillary GIS layers that flag geographic features normally associated with human development (i.e., road network, towns, mining locations, etc.); and published documents and web sites describing specific land use changes occurring in the North.
 Because clear-sky Landsat scenes were not always available for the exact years of interest (i.e., 1985, 1990, or 2000), an adjustment was applied to the areas where the imagery dates fell outside of the targeted 1985–1990 and 1990–2000 intervals. This adjustment involved dividing the LUC area by the number of years spanning each pair of Landsat scenes, then assigning only that proportion falling within the pre-1990 and 1990–2000 intervals. For example, if a 10 ha LUC event was recorded between 1989 and 2001, the annualized area would be 10 ha/12, or 0.83 ha. One year of LUC (0.83 ha) would then be assigned to the pre-1900 interval, while a 10-year proportion of LUC (8.3 ha) would be assigned to the 1990–2000 interval.
A2. Approach for Mapping Aboveground Biomass for 2000
 Chen et al.  investigate the relationships between field measured aboveground biomass and optical and microwave remote sensing data. The results showed that the synergistic use of Landsat 7/ETM+ and JERS-1/SAR data can best describe the measured aboveground biomass, and significantly increase the accuracy of estimating aboveground biomass. Using measurements from the Dempster Highway study area, we established relationship between aboveground biomass Ba and the combination of Landsat band 4, 5, and JERS-1 backscatter, given by
with r2 = 0.72 and standard error of estimation (SEE) = 0.78. The validation of equation (A1) using measurements from Yellowknife/Lupin study area indicated that question 2 had good transferability. The total number of sites of the two study areas = 75. Because the logarithmic equations could introduce a systematic bias when used for back calculating biomass, it has now become fairly widely recognized that a correction factor is necessary to counteract this bias [Sprugel, 1983]. The correction factor (CF) can be calculated by using the formula:
For equation (A1), the CF = 1.35.
 The circa 2000 aboveground biomass maps for main LUC areas were produced by applying equations (A1) and (A2) to the corresponding 2000 summer Landsat mosaics and the 1998 summer JERS mosaics.
A3. Approach for Estimating Aboveground Biomass Carbon Stock Change due to LUC During 1985–1990 and 1990–2000
 The right-hand side of Figure 2 shows the flowchart for estimating aboveground biomass carbon stock change caused by LUC. For LUC from natural vegetation to settlements in Canada's north, such as mining, commercial and residential area development, airport and transportation development, we can reasonably assume that the aboveground biomass is zero after the land use change. Therefore, the aboveground biomass change due to a LUC from natural vegetation to settlements is zero minus the pre-LUC aboveground biomass in the LUC area, Ba,p (t ha−1):
Since the conversion factor from oven-dry biomass to carbon is about 0.5, ranging from 0.45 to 0.55 according to vegetation species, age, formation, and community structure [Olson et al., 1983], the corresponding carbon stock change due to LUC, ΔC, can be calculated by
To convert carbon stock change to CO2 emission, we multiplied ΔC by 44/12. In this study, a negative ΔC represents carbon release from the land to the atmosphere, and a positive ΔC represents carbon removals from the atmosphere into the land.
 Conversely, for LUC from settlements to natural vegetation (e.g., previous mining site renegotiation), Ba,p would be zero, and the post-LUC aboveground biomass, Ba,a (t ha-1), can be estimated from remote sensing data using equations (A1) and (A2). The aboveground biomass change due to a LUC from settlements to natural vegetation = the post-LUC aboveground biomass in the LUC area, Ba,a minus 0:
and the corresponding carbon stock change due to LUC, ΔC, is given by
Because the majority of LUC in Canada's north has been from natural vegetation to settlement, we focused on the calculation of Ba,p. Calculation of Ba,p using equations (A1) and (A2) needs both Landsat and JERS-1 backscatter data. Because the JERS-1 backscatter data are only available for the summer of 1998, an alternative “replacement” method has to be used to estimate the value of Ba,p.
 This “replacement” method assumes Ba,p = the average aboveground biomass of surrounding areas that have the same land class (i.e., sparse woodlands, shrub, or grass) as the LUC area prior to its land use change event. The underlying assumption of this method is that there should be little change in vegetation biomass without human-induced or natural disturbances, and that on average the values of biomass in a small, adjacent area vary little for the same land cover class. It would probably be true in Canada arctic and subarctic region because the vegetation grows very slowly. Since vegetation along the edge of a land use change area can be affected by LUC activities, the surrounding area is found outside of the buffering zone as shown in Figure 3. The land use change areas come in two shapes: polygon and line. Since a line may extend to several hundred kilometers, each pixel in the line will be treated as individual land use change area. Here the surrounding area of a land use change area is defined as a buffering zone between 100 and N meters (usually N > 200 m) from the land use change area. The following steps describe the details of the method for estimating Ba,p for a LUC area.
 1. Identify the land cover type, c, for every 30 m by 30 m Landsat pixel within a LUC area, C, on the pre-LUC land cover map, where c can be sparse woodlands, shrub, or grass.
 2. Calculate the area of land cover class c within the LUC area, Ac.
 3. Search pixels of land cover type c in the surrounding area of the land use area until the number of pixels for this land cover class, nc, reaches a criteria (e.g., 200 pixels). If not enough pixels are found, enlarge the surrounding area and search again.
 4. Extract biomass value for each pixel of land cover class c in the surrounding area, Bc,i, from the circa 2000 aboveground biomass map.
 5. Estimate Ba,p for the LUC area C by using following equation: