Areal changes of land ecosystems in the Alaskan Yukon River Basin from 1984 to 2008

Multivariate alteration detection (MAD) and Bayesian inference (BI) methods are used to analyze land cover changes with Landsat images for the Alaskan Yukon River Basin from 1984 to 2008. The US Geological Survey National Land Cover Database 2001 (NLCD 2001) is treated as reference information to detect the changes. It is found that the regional land cover change has three general trends with various potential causes during the study period: (1) forests decreased mainly due to wildfire, (2) the closed water bodies were shrinking possibly due to permafrost degradation if water drains well in discontinuous permafrost regions, (3) shrubs had expanded and a large portion of grassland was converted into shrubland likely due to forest fire and warming. The uncertainty of this analysis may mainly arise from image acquisition date differences and illumination angles and remaining cloud contamination to the images. This study provides a method to analyze land cover changes with Landsat data for other regions. The developed land cover data should help future understanding of permafrost dynamics, biogeochemistry, hydrology and regional climate in the region.


Introduction
Global surface temperature has increased at a rate of 0.13 ± 0.03 • C/decade for the past 50 years [1]. This warming trend is more evident in high northern latitudes [2]. In particular, the climate in Alaska switched its state abruptly from cool and moist to hot and dry in the summer of 1976 while no statistically significant change in annual precipitation was reported from 1951 to 2001 [3]. This warm and dry shift in the climate regime potentially altered land cover [4][5][6] including the shrinkage of wetlands, drying and succeeding to upland habitat [7], changing of water bodies [8][9][10] and expanding of shrubs [11,12] in the region.
Analysis of land cover changes is important for evaluating the transition of ecosystems, the carbon cycle and the water and energy balance in response to climate changes in the subarctic region. The change of water bodies and associated wetland areas will in turn affect the carbon balance, including methane dynamics in the sub-arctic [13,14]. To date, there is a lack of studies examining land cover changes in a high spatial resolution manner in the sub-arctic. Satellite sensors with a 1 km or coarser resolution are not suitable for detecting the dynamics of relatively small water bodies and new sensors with finer resolutions lack long-time records. In contrast, the Landsat satellites provide a longer temporal record of land observations and Landsat data have been used to reconstruct the history of the Earth's land surface back to 1972 for various regions [15].
The Yukon River Basin is the fourth largest basin in North America with a drainage area of 832 700 km 2 , of which 508 900 km 2 is in Alaska (figure 1). It has large variations in topography: low rolling hills and low mountains account for 61% of the total area, plains and lowlands account for 20% and around 19% is occupied by high rugged mountains. The region has permafrost underneath, which can be categorized into five types [16]: continuous permafrost with high, medium and low ice content, each accounting for 1.2%, 17.2% and 11.6% of the permafrost area, respectively, and discontinuous permafrost with medium and low ice content accounting for 36.3% and 33.6% of the permafrost area, respectively. The analysis of land cover change in the region will help our understanding of the carbon and water cycling. However, to date, there is a lack of efforts to analyze the land cover change in the region. Here we take advantage of the available Landsat images to investigate the land cover change in the Alaskan Yukon River Basin from 1984 to 2008. We further explore the possible causes for the identified changes.

Overview
We first obtained Landsat images from 1984 to 2008 and National Land Cover Database 2001 (NLCD 2001) data for the region [17]. Second, we removed the effects of terrain, cloud and associated shadow from the Landsat images. Third, we used the multivariate alteration detection (MAD) [18][19][20] and a Bayesian inference (BI) method [21] to identify the areal change using the NLCD 2001 data as reference. Fourth, we corrected the possible classification biases using the probabilistic label relaxation (PLR) method [22]. Finally, we analyzed the potential causes for the areal changes of major land cover types. More detailed method information is provided as supplementary material (see figure S1 available at stacks.iop.org/ERL/6/034012/mmedia for the overall flowchart of the land cover mapping method).

Satellite data acquisition and pre-processing
Landsat TM and ETM+ images for the four periods of 1984- 9, 1990-5, 1996-2001 and 2003-8 in the Yukon River Basin were obtained (http://glovis.usgs.gov). Each period consisted of 26-31 scenes and all images were selected for the period between 15 June and 30 August to minimize errors due to seasonal and annual phenological cycles (see the supplementary material for the data used, available at stacks.iop.org/ERL/6/034012/ mmedia). Each period consisted of 26-31 scenes (total of 155) selected for peak vegetation cover and minimal cloud cover to minimize errors due to seasonal and annual phenological cycles. Two thirds of the scenes were between 15 June and 30 August, with the dates of the additional 52 scenes ranging from 16 May to 24 September. The cloud cover category was 0% for most scenes, with 12 scenes with 1-9% cloud cover, 14 scenes with 10% and 4 scenes with 20%. The images acquired during or near significant rainfall events were excluded based on meteorological precipitation data [23].
The automatic cloud cover assessment (ACCA) algorithm was used to detect clouds over the images [25]. Further, a combination of geometric and optical constraints was used to detect cloud shadows [26]. Specifically, we estimated the position of the cloud shadow from shadow length and direction. We acquired the cloud shadow length for each scene and determined the direction of the cloud shadows based on the solar azimuth angle. Next we used the semi-empirical cosine correction (C-correction) method to correct the images for the effects of local solar illumination [27]. The required digital elevation model (DEM) for processing the Landsat data was extracted from the Global, 1 arc second (approximately 30 m) digital elevation model (GDEM) [24].

Change detection and classification
With the corrected Landsat data, the land cover change for periods other than 2001-3 was detected based on NLCD 2001 data [17]. In this study, images acquired during 2001-3 were used as reference image and it was assumed they have the same land cover distribution as NLCD 2001. Using the MAD method, a pair of two images at a given time was spectrally compared. One was the reference image during the period of 2001-3 and the other image was to be analyzed with respect to land cover changes. First, all raw digital number (DN) records of these images were converted into top-of-atmosphere (TOA) reflectance data [28]. The six spectral channels (without the thermal band) in the two images were then linearly transformed into two variates whose transformation coefficients were determined by minimizing the positive correlation between the two variates. The change detection analysis was then applied to the difference between two MAD variates. Note that the bias due to the sensors or atmospheric factors was also minimized in the transformation, thus the difference between the two MAD variates only provided the information on the land cover change between the two images. For example, the Landsat ETM+ color composite scenes of band5 (red), band4 (green) and band3 (blue) acquired over a lake in the Yukon River Basin in August 1999 and July 2001, respectively were analyzed (figure 2). The detection results were derived from the first MAD component (MAD1) and the first two components MAD1 and MAD2, respectively. Here we only used MAD1 in our analysis. The difference between the two MAD variates was approximated using three Gaussian components that correspond to negative change, no change and positive change, respectively. If there is no change, the MAD variates are near zero. Negative and positive values of MAD variates indicate changes. The BI method was used to determine if there was a change in MAD variates. Detailed information on the MAD and BI methods is provided in the supplementary material (available at stacks.iop.org/ERL/6/ 034012/mmedia).
Next, we randomly selected 25 000 no-change pixels to train classifiers for land cover types for each scene. We combined the NLCD classification system into eight classes as the final results (see the supplementary material for the combination rule, available at stacks.iop.org/ERL/6/034012/ mmedia). Due to the linear treatment of the atmospheric effect and the resulting uncertainties in the transformation coefficients and the use of the pixel-oriented classifier, there was some 'salt-and-pepper' appearance in the resulting thematic map. The PLR method [22] was then used to improve the spatial coherence. To reduce the effect of flood and the seasonal change of river channels, we delineated streams that were longer than 1 km in the Yukon River Basin based on the National Hydrograph Dataset (NHD, http://nhd.usgs.gov/). The areas within a distance of 150 m from the streams were ignored.

Results and discussion
The land cover maps for the periods 1984-9 and 2003-8 and corresponding change detection results are shown in    or insect damage might be the most likely causes for the change. Similarly, when we overlay the fire polygons of 1991 and 1992 on the forest disappearing regions between 1984-9 and 1990-5 (not shown here), we find that the wildfire records match the areal change during these periods.

Change of water bodies
Water surface area decreased by 12.6% from the period 1984-9 to 2003-8 in this region (table 1). The water surface refers to the opened and closed lakes and ponds that lack inlets and outlets, but does not include streams. Since the open water bodies are steadily affected by upstream flows through hydrological networks, we only studied the distribution of the closed water bodies during the study periods. The NHD data, which provide the detailed stream information in the Yukon River Basin, were used to extract the closed water bodies. The lakes and ponds intersected with streams were excluded from further analysis. We also omitted the water bodies whose area was less than 3600 m 2 . Note that the study periods of 1984-9 and 2003-8 have relatively better image quality in terms of cloud fraction and the Sun's elevation angle. Although the cloud removal and terrain correction was carried out, some thin clouds still existed and the effectiveness of the terrain correction was not perfect. As a result, some shadows on images still exist and are wrongly classified as water bodies. Consequently, our analysis of the dynamics of the closed water bodies is only focused on these two periods.
The total number of closed water bodies (larger than 3600 m 2 ) declined from 168 005 during 1984-9 to 160 965 during 2003-8. The area of the closed water bodies slightly decreased from 764 851 to 764 347 ha during the 25-year period. We conclude that the Yukon River Basin has a general decline in closed water bodies. As expected, the closed water bodies show a smaller change (table 2) in comparison to the dynamics of all water bodies (table 1). Most of the region has permafrost beneath the topsoil. Thus, we incorporated the permafrost distribution [16] into our analysis of the spatial pattern of the closed water bodies' changes. The sporadic permafrost was not considered due to its tiny fraction. The closed water body dynamics varied significantly depending on the different permafrost types (table 2). The general trends are: (1) the closed water bodies increased in both their area and count in continuous permafrost-dominated areas and (2) the closed water bodies underlaid by discontinuous permafrost showed an opposite trend in their count and area in comparison to the continuous permafrost-dominated areas. These results are consistent with a previous study in western Siberia [8]. Fresh water does not infiltrate the continuous permafrost area due to an impermeable layer of permafrost. On the other hand, water above discontinuous permafrost can easily infiltrate the subsurface of the hydrological system. From table 2, we assume that the ice content of permafrost is the fundamental reason for the change of the closed water bodies. Specifically, the closed water bodies in continuous, low ice content permafrost regions may show a similar change trend to those in discontinuous permafrost regions. On the other hand, the closed water bodies in discontinuous, medium ice content permafrost regions could also show dynamical trends like those in the continuous permafrost regions. The key is whether fresh water can find pathways to join the groundwater system. If yes, the closed water bodies shrink; if not, they expand. In the Yukon River Basin, 54% of the area is underlaid by permafrost with high or medium ice content while the rest has low ice content. The increase of the closed water bodies on the former is offset by the decrease of the latter. The growing season (May-September) air temperature in the region persistently increased during the study period and permafrost degradation in this region has been reported by a previous study [30], while precipitation fluctuated [23]. Thus, the thawing permafrost very likely plays an important role in the changes of the closed water bodies.

Changes of shrubland, grassland, and wetland
Shrubland (shrub/scrub and dwarf scrub) increased from 37.2% to 38% from 1984-9 to 2003-8. A large forest area was converted to shrubland ( figure 6(a)). Most of the area has  fire history, suggesting that forest fire is an important factor in shrub expansion. Meanwhile, some wetlands were drying and succeeded by shrubs and some barren land and grassland was also converted into shrublands. Grassland/cropland in the Yukon River Basin showed a very small decrease during the 25-year period (table 1). From 1984-9 to 2003-8, some grasslands were invaded by shrub/scrub ( figure 6(b)). This is consistent with other studies in this region [12,31]. Further, about 46% of grassland loss area was taken by shrub/scrub, which might be more likely due to succession. Warming may be another cause. At the same time, some grassland/cropland land covers have been converted into wetlands. We find that 64% of these new wetlands are wetlands that usually have a high variability due to flood events.
Wetland variation showed no consistent pattern (table 2). The areal change of wetlands might be greatly affected by short-term precipitation events. With warming, wetlands could expand in the areas dominated by the permafrost having high or medium ice content where the closed water bodies are increasing. In contrast, wetlands may also shrink due to the decline of the closed water bodies in the low ice content permafrost-dominated areas. Landsat TM or ETM image data are sufficient to separate wetlands from uplands [32], but errors occur in separating wetlands from forests [33]. Thus, wetland changes in the region are subject to further investigation.

Summary
The NLCD 2001 data were used as reference information for mapping the land cover change for the four periods from 1984 to 2008. We found that the Yukon River Basin had the following land cover change patterns during the study period: (1) forests were decreasing mainly due to wildfire, (2) the closed water bodies were shrinking possibly due to permafrost degradation if water drains well in discontinuous permafrost regions, (3) shrubs had expanded and also a large portion of grassland was converted into shrubland likely due to forest fire and warming. The results are likely to have a great uncertainty mainly due to acquisition date differences and illumination angles and remaining clouds contamination in the images we used. For instance, the scan line corrector failure on ETM+ made candidate images even less available for the period 2003-8. Consequently, some images with large acquisition date differences from their reference data were used in order to cover the whole Yukon River Basin in our analysis, which will result in uncertainty. Next, we will use the land cover change data and associated uncertainty information to study the carbon and water dynamics during the last 25 years with biogeochemistry modeling approaches for the region (e.g., [13,14]).