The Coal River Watershed is a 230,755 ha Hydrologic Unit Code (HUC) 8 headwater watershed completely within West Virginia, USA. The Coal River Watershed exists within the Appalachian Plateau physiogeographic province, a dissected, westward-tilted plateau dominated by Pennsylvanian bedrock. Pennsylvanian stratigraphy is characterized as cyclic sequences of sandstone, shale, clay, coal, and limestone [18]. The terrain is dissected by a dendritic stream network and shows fine texture with moderate to strong local relief. In comparison to the northern Appalachian Plateau, the Southern Coalfields is generally more rugged due to resistant strata [19]. The terrain analysis using landforms suggests that this terrain is naturally dominated by steep slopes.

According to LULC estimates derived from aerial photography, 8.8% of this watershed was disturbed by active surface mining or mine reclamation in 2009. It should be noted that this estimate does not take into account historical mining areas which have since been reforested. Figure 1 shows the watershed location within West Virginia. This watershed was selected as a case study because it is heavily impacted by mining, there has been continued mining and reclamation between 2003 and 2010, and because LiDAR data were available for the extent.

The LiDAR data, representing recent topographic conditions, were collected between April 9^th, 2010 and April 18^th, 2010 at a flight height of 1524 meters above ground level (AGL), a pulse frequency of 70 kHz, a scan frequency of 35 Hz, and a scan angle (full field of view) of 36˚. The swaths were flown with a 30% overlap, at an average speed of 250 km/hr, and with an average width of 979 m. An Optech ALTM 3100 C sensor was used to collect the data. The scan and flight specifications were selected to support a 0.7 m contour interval and a 1 m nominal ground post spacing. Ground points were filtered from the all return data to produce ground point files in LAS 1.2 format.

An interpolated raster surface was created from point data using ArcMap 10, and average point spacing of 0.01 m. Inverse distance weighting (IDW) was used to interpolate a raster surface since sample points were in an evenly distributed pattern. This process resulted in a 1 m resolution, floating point elevation dataset.

Landform changes were assessed in filled and cut or excavated areas resulting from mining and reclamation between 2003 and 2010. The 2010 LiDAR-derived DEM was compared to a 2003 photogrammetrically-derived DEM representing 2003 conditions. The 2003 DEM data were provided by the West Virginia Statewide Addressing and Mapping Board (SAMB) who contracted BAE Systems ADR to create a stereo photogrammetric-derived digital terrain model (DTM) from statewide, spring, leaf-off, 1:4800 scale aerial photography. The DEM was created in compliance with National Dataset standards (1/9^th arc second) and produced from mass points and breaklines. This dataset supports a vertical accuracy of +/− 3.048 m, root mean square error (RMSE) and has a cell size of 3 m [20].

In order to match the resolution of the SAMB DEM, we resampled the LiDAR-derived raster to 3 m resolution using bilinear interpolation and snapped the grid to match the extent of the SAMB DEM. This process assured that the raster grids were completely aligned. This process resulted in two 3 m DEMs representing 2003 and 2010 conditions that could be compared. A vertical data transformation was unnecessary since both DEMs had a vertical reference of NAVD88 orthometric.

In order to detect systematic error between the 2003 and 2010 DEM data, a total of 281 points were collected in the field throughout the Coal River Watershed on flat, paved surfaces using Pacific Crest realtime kinematic (RTK) survey equipment. The elevation measurements from the 2003 and 2010 elevation raster data were obtained at these locations. The measurements from each DEM were then compared. The mean difference was −0.4 m with a maximum difference of 0.8 m and a minimum difference of −2.0 m. Based on this analysis we did not correct for systematic difference between DEMs.

Once DEMs were obtained, processed, and prepared, they were subtracted using Raster Calculator within the Spatial Analyst Extension of ArcMap [21]. This produced a grid of elevation differences throughout the watershed. Negative values indicated potential cuts or excavations while positive values indicated potential fills. However, difference could simply have resulted from errors in the DEMs or in methodology. As a result, once this elevation difference model was calculated, it was necessary to determine a tolerance or threshold that would constitute true change and not simply error or noise between the digital elevation datasets. The photogrammetric DEM had an error tolerance (RMSE) of +/− 3.048 m while the LiDAR DEM had and error tolerance of only 15 cm (0.15 m). An equation suggested by [22] was used to estimate this threshold. The RMSE of each dataset was squared, the results were summed, and the square root of the sum was then taken. The result was then multiplied by three to determine a cut off representing values that were greater than three standard deviations from the mean. This method assumes a Gaussian distribution. Elevation change measurements outside of this range were considered true elevation differences and not a result of error or noise. Reference [15] used the same methodology to derive an error tolerance for the analysis conducted in Perry County, Kentucky, USA.

According to this method, it is not certain whether differences less than +/− 9.2 m represented true topographic change. As a result, the elevation difference grid was reclassified so that values between −10 m and +10 m were considered no change, error between the DEMs, or noise. Pixels with values less than −10 m were considered potential cuts while pixels with values greater than 10 m were considered potential fills. This process produced a cut and fill mask.

A LULC dataset was created for the region from 2009 National Agriculture Imagery Program (NAIP) orthophotography. Optimally, imagery collected at the time of the LiDAR collection would have been utilized; however, such data were not available. The imagery was classified using an object-based feature extraction methodology, augmented with GIS decision rules and manual digitizing. Forest, grass, and barren land cover were extracted from the raw imagery.

Combining these data with mine permit boundaries from the West Virginia Department of Environmental Protection (WVDEP) made it possible to delineate barren and grassland land cover in mine permits. Grasslands in permits represent potential reclamation while barren areas represent areas of active mining or areas that have not yet been reclaimed. The extents of valley fill faces and slurry impoundments were digitized. According to an error assessment based on manual aerial photograph interpretation at 100 randomly selected points, the LULC dataset had an overall accuracy of 94% (K_Hat of 93%).

It was possible to further remove erroneous pixels as potential cut or fills by utilizing additional data, such as the high resolution land cover. For example, potential cut or fill pixels could be removed if they were outside of disturbance resulting from surface mining. The potential cut or fill pixels that existed within areas of mine disturbance (barren, grasslands, or valley fill faces) were considered. Slurry impoundments were excluded because of errors associated with water level. To complete this analysis, it was necessary to convert the raster data to polygons that represented contiguous areas of cut or fill. No smoothing or simplification was applied.

Small areas of topographic change could also be removed as error or noise. Thresholds were determined from the hand digitized valley fill faces. The average size of these faces was 73,257 m² (7.3 ha) with a standard deviation of 67,013 m² (6.7 ha). Based from this distribution, any contiguous area of cut or fill that was smaller than 60,000 m² (6 ha) was considered noise and removed. Although this threshold was somewhat arbitrary, it was selected since this study focused on detecting large expanses of excavation and fills.

The result of this analysis was a vector layer of potential cut or fill extents in which topographic change had occurred between 2003 and 2010. The extents were used to define areas of potential terrain change for analysis and comparison of the distributions of landforms. Although there was error associated with this methodology, it was adequate for this analysis since we were not interested in conducting an assessment of absolute elevation change or volumetric change but in defining areas in which to compare categorical terrain data and their impact on forest communities.

The DEM data were used to classify the terrain into the following landforms: cliff, steep slope, slope crest, upper slope, flat summit, sideslope, cove, dry flat, moist flat, wet flat, and slope bottom. The methodology described by [17] provides a means to utilize DEM data to classify the landscape into different units of ecological significance. DEM derivatives including slope in degrees, a hydrologically filled DEM, flow direction, and flow accumulation were created using ArcMap 10. The slope and flow accumulation grids were utilized to calculate a moisture index. The moisture index is a relative measure of the moisture of a specific cell, and it assumes that the moisture level is a function of how much water flows into the cell, predicted by flow accumulation, and how fast the water can flow out, described by slope [20-25].

Landscape position was calculated following [26] to divide the landscape into the following categories: ridge, wide ridge, slope/flat, slope/cove. The approach is based on a local neighborhood analysis of elevation values in which a cell is compared to the mean elevation of all values within 3 by 3 windows. Landscape position was combined with slope and moisture index data to derive the final landform classes. This procedure was conducted for both the 2003 and the 2010 DEMs. The comparisons of the derived landforms included:

•      Areas filled between 2003 and 2010 (mining related);

•      Areas cut or excavated between 2003 and 2010 (mining related);

•      Areas of mine reclamation in 2009, areas of mining disturbance or reclamation in 2009, and areas in WVDEP mine permits but not disturbed (forested in 2009).

This enabled the comparison of post-mining landform categories with pre-mining conditions.