retroactive comparison of operator-designed and computer-generated skid-trail networks on steep terrain

Aim of the study : Quantify potential economic benefits of implementing computer-generated skid-trail networks over the tradi­ tional operator-designed skid-trail networks on steep terrain ground-based forest operations. Area of study : A 132-ha harvest operation conducted at the University of Kentucky’s Robinson Forest in eastern Kentucky, USA. Materials and methods : We compared computer-generated skid-trail network with an operator-designed network for a 132-ha har­ vest. Using equipment mounted GPS data and a digital elevation model (DEM), we identified the original operator-designed skid-trail network. Pre-harvest conditions were replicated by re-contouring terrain slopes over skid-trails to simulate the natural topography and by spatially distributing the harvestable volume based on pre-harvest inventories and timber harvest records. An optimized skid-trail network was designed using these pre-harvest conditions and compared to the original, operator-designed network. Main results : The computer-generated network length was slightly longer than the operator-designed network (53.7 km vs. 51.7 km). This also resulted in a slightly longer average skidding distance (0.71 km vs. 0.66 km) and higher total harvesting costs (5.1 $ ton -1 vs. 4.8 $ ton -1 ). However, skidding costs of the computer-generated network were slightly lower (4.2 $ ton -1 vs. 4.3 $ ton -1 ). When comparing only major skid-trails, those with ≥ 20 machine passes, the computer-generated skid-trail network was 28% shorter than the operator network (9.4 km vs. 13.1 km). Research highlights : This assessment offers evidence that computer-generated networks could be used to generate efficient skid­ trails, help determine skidding costs, and assess further potential economic and environmental benefits.

introduction Timber harvesting operations on gentle terrain are performed with ground-based systems using skidders or forwarders, while cable systems are recommended on steeper terrain (Kellogg et al., 1992). However, in many parts of the eastern US such as the Cumberland Plateau region of Kentucky typified by relatively steep, highly dissected terrain with short distances, the effective use of cable systems has been difficult to establish, and ground-based operations are com mon. As opposed to gentle terrain areas where skid ders can travel relatively unrestricted, steeper areas require constructed skid-trails to facilitate cost-effec tive and safe operations. Consequently, efficiently locating skid-trails becomes crucial as they directly impact skidding and skid-trail construction costs. Typically, skid-trail networks are designed manually by managers using vegetation and terrain character istics but more often are constructed on-the-fly by a bulldozer operator without careful planning. Typi cally, bulldozer operators start building skid-trails 2 either along ridge lines or near stream corridors and subsequently along contour lines. This results in relatively parallel skid-trials, spaced between 45 m and 75 m depending on harvest machinery and crew resources to facilitate reaching all harvestable volume between skid trails.
The heavy traffic of harvesting equipment along skid-trails has also been reported to cause significant soil disturbances that can lead to erosion and com paction (Croke et al., 2001;Williamson & Neilsen, 2000), a shift in vegetation composition (Avon et al., 2013;Buckley et al., 2003), and loss of vegetation productivity (Lockaby & Vidrine, 1984). Best man agement practices including disking and seeding, subsoiling, re-contouring, and installing water bars are often recommended to ameliorate soil distur bances (Conrad et al., 2012). However, these prac tices carry additional costs ranging from 500 $ to 8,000 $ ha -1 that might cause significant economic impacts on timber harvesting operations (Soman et al., 2019;Sawyer et al., 2012). The effort and costs used to ameliorate soil disturbance is partially governed by, and positively related to the traffic level. Reducing the length of high-traffic skid-trails can help alleviate administrative costs and thus des ignated skid-trails is typically recommended to also reduce these soil disturbance (Garland, 1983;Han et al., 2006).
There are only a few models to automate the design of optimized skid-trail networks. Halleux and Greene (2003) developed an automated approach to evaluate alternative networks but assumes flat terrain and evenly distributed volume. Gumus & Turk (2016) developed an approach to optimize the design but is also applicable only for flat terrain. Contreras et al. (2016) developed a computerized model to generate an optimized skid-trail network that minimizes skid ding and skid-trail construction costs based on terrain, volume distribution, and extraction locations. Despite these developed models, there has been no formal comparison between field implementation of comput er-generated and operator-designed skid-trail networks to quantify potential economic benefits. One of the main reasons for the lack of these studies is the re quired coordination and collaboration with forest companies and logging contractors. Other reasons are the logging contractors' unwillingness to change tra dition, perceived costs associated with tasks such as flagging skid-trails before construction, and an in herit distrust and misunderstanding of computer generated resources.
In this study, we retroactively compared an oper ator-designed skid-trail network for a harvest op eration conducted in eastern KY, USA in 2008 with the optimized computer-generated skid-trail network using the Contreras et al. (2016) model. This work presents a novel attempt to quantify potential eco nomic benefits of computer-generated skid-trail networks, which can facilitate future more compre hensive ground comparisons and evaluation of model applicability.

Study area
The study site was in the University of Kentucky's Robinson Forest (lat. 37.47° N, long. -84.24° W), lo cated within the Northern Cumberland Plateau region in eastern Kentucky. The landscape is deeply dissected with steep slopes, and the forest overstory is primarily composed of oak (Quercus spp.), yellow-poplar (Liriodendron tulipifera L.), and hickory (Carya spp.). For the study, we focused on three watersheds, totaling 132 ha, harvested in May 2008 to August 2009. A deferment harvest with a target residual basal area of 3.4 m 2 ha -1 was performed resulting in the removal of 16,164 tons of merchantable products. Full-benched skid-trails were constructed mostly along contours by the operators of three bulldozers: John Deere 650, John Deere 700, and John Deere 850. On accessible slopes below 30%, a Timbco 445 EXL feller-buncher was used to fell, top, and delimb trees. On steeper slopes the feller-buncher was restricted to the skid-trail and operated within reach of the boom. Trees beyond the reach of the boom were manually processed and merchantable length stems were winched to skid-trails by a bulldozer. Log piles created by the feller-buncher and the bulldozer were skidded to three landings by Caterpillar 545 grap ple skidders. Landings were located on ridgetops result ing in uphill skidding throughout much of the har vested area.

Simulating pre-harvest conditions
A high-density (~25 pt m -2 ) LiDAR dataset acquired in the summer of 2013 was used to create a high resolution digital elevation model (DEM) of the study area. While the DEM was created from data collected 5 years after the harvest, the remnant skid-trail net work was clearly visible. To ensure a fair comparison with the computerized skid-trail model, we removed these terrain disturbances and created a DEM that mimicked the terrain prior to the harvest for input into the computerized skid-trail model program. Using the high-resolution DEM, aerial photos, and GPS data 3 Comparing operator-designed and computer-generated skid-trail networks collected from units mounted on the harvesting equip ment, the operator-designed skid-trail network was identified, and each skid-trail segment was digitized as a line through the center of each skid-trail (Fig. 1a). A 6-m buffer centered on the digitized skid-trail net work was applied to encompass the entire area dis turbed by skid-trail construction. Elevation data from the DEM cells within the buffer were removed and a routine was developed to fill the vacant elevation data. The elevation of a given DEM cell within the buffer was calculated as the inverse distance weighted aver age of the elevation of the closest DEM cell along eight transects starting from north and generated every 45 degrees.
Harvested volume was spatially distributed across the study area using pre-harvest inventory data con sisting of a systematic grid of 186 points. The inven tory used a nested variable point sampling for trees with diameter at breast height larger than 33 cm, and the variable point sampling with diameter obviation method described in Beers (1964) for smaller trees. The inventory only recorded trees that were marked for harvest. It was assumed that harvested volume estimates per sample point were representative of the volume distribution across the study area. Then, har vested volume per ha across the watersheds was esti mated by interpolating the volume estimates from the sample points. The interpolation procedure used the inverse distance weighted method to create a 1-m distribution raster with the percentage of the total extrapolated volume for each cell covering study area. To ensure that the recreated pre-harvest volume was equivalent to the actual harvested volume, sale tickets from the harvest were used to calculate the exact volume extracted from each watershed. This total volume was then distributed according to the distribu tion raster.

computerized skid-trail network model
The model presented in Contreras et al. (2016) was used to develop the computer-generated skid-trail network. The model creates an optimized skid-trail network based on a DEM, volume distribution, skid der maximum loading capacity (MLC), obstacles within the harvesting area, and costs of skid-trail construction and skidding. Based on the volume dis tribution by cell and the skidder's MLC, the model uses a log-bunching routine to identify the location of log-piles. In the volume raster, the routine identi fies the first accessible cell with volume and adds the volume to the first log-pile. If the volume is less than the MLC, the routine searches the neighboring cells for additional volume. If present, the volume is added, and the cell is assigned to the log-pile. The search window continues to expand to add addi tional volume and assign the associated cells to the log-pile until the pile volume equals the MLC. Once this target volume is achieved, the log-pile location is established in the center of the search window area. The model then identifies the next unassigned cell with available volume, adds additional volume from an expanding search window, assigns the cells to the next log-piles, and when the volume meets the target MLC the center of the search window area is assigned as the location of this next-log pile. The process con tinues until all cells with volume are assigned to a log-pile.
The model creates a network of feasible skid-trail segments formed by a set of vertices regularly spaced throughout the study area and links connecting adjacent vertices. Vertices represent the center of DEM cells, log-pile locations, and landing locations. Links repre sent skid-trail segments between adjacent vertices. In the model, each vertex was connected to eight adjacent vertices spaced every 6.4 m (20 ft) over trafficable areas with gradient and side slopes below user-defined limits for skidding. Skidding costs for skid-trail seg ments were calculated based on skidder rental rate and cycle time where the cycle time for uphill and downhill links were determined using the following equations from Contreras and Chung (2007): where CT ds is the cycle time (min) for downhill skid ding, CT us the cycle time (min) for uphill skidding, and D the slope distance (m) along the network connecting a log-pile and the landing. Cycle time was used to calculate skidding cost as follows: where PSC i is the skidding cost ($) for the i th log pile, CT i round trip skidder cycle time (min) for the i th log-pile, and RR the hourly rental rate for the skidder ($).
As model inputs, slope limitations for feasible skid trail segments (links) were set to not surpass 45% gradient slope and 100% side slope. Skidder rental rate was set at 120 $ SMH -1 (US Forest Service, 2011) and MLC was set as 10 ton based on cycle volume observations for similar harvest operations near the study site (Bowker, 2013). To estimate skid-trail con struction cost, the same rental rate associated with skidding was used, 120 $ hr -1 (US Forest Service, 4 2011). Construction time was obtained from the GPS positional data with timestamps mounted on the three bulldozers and collected during the original harvest (Bowker, 2013). Using construction time and average terrain side slope along each skid-trail section, we found a 30% decrease in construction time within each 10% increase in terrain slope. Applying this relation ship to the average slope and average time of the original harvest, we estimated construction time for each skid-trail segment based on slope distance and terrain side slope. Streamside management zones in the original harvest were identified and considered inaccessible in the model (Fig. 1b). Lastly, NET WORK 2000 (Chung and Sessions, 2003) was used to find the optimal skid-trail network considering variable (skidding) and fixed (skid-trail construction) costs and connecting each log-pile to the three land ings at minimum total costs.

comparison of skid-trail networks
Although constructed skid-trails of the operator designed network could be easily identified, the location of skid-trails that were not constructed and used to access and pick up individual log-piles were unknown. Thus, to be consistent with the computer ized model inputs, the same log-pile locations gener ated by the log-bunching routine were assumed to represent the locations of the log-piles in the original harvest. These log-piles were then linked to the iden tified operator-designed skid-trails with Euclidean distance lines with no restrictions on terrain slope. The operator-designed skid-trails were divided into 3.05 m segments, for which skidding and construc tion costs were calculated following the same pro cedures as in the computerized model. Routes and skidding cycle times for each log-pile were deter mined assuming the shortest distance along the op erator-designed skid-trail network to the nearest landing. Then, information per log-pile (i.e., skidding distance and costs) was determined, as was informa tion for the entire study area (i.e., skidding cost, skid-trail construction cost, total harvesting cost and skid-trail length). These were calculated and com pared with the information from the computer-gen erated skid-trail network.
The potential economic benefit of optimizing the location of skid-trails is proportional to traffic level. Thus, we compared the total length of both skid-trail network for segments with increasing levels of ma chine passes. Typically, in moderately steep areas such as our study area, skid-trails need to be constructed even across low-volume areas to be able to reach log piles. Therefore, we also focused comparisons on major skid-trails.

results and discussion
The total harvest volume represented by the vol ume distribution data in the harvest area (Fig. 1c) was 16,021 tons, from which the log-bunching rou tine identified 1,667 log-piles (Fig. 1d). The average log-pile volume was 9.6 ton, which was near the skidder maximum capacity set at 10 ton. The oper ator-designed network presented the typical parallel pattern with an average spacing of 56 m (Fig. 1a). Figure 2a shows the complete operator-designed skid-trail network after connecting all log-piles to the closest constructed skid-trails. The number of loaded machine passes ranged from one, for seg ments connecting log-piles to constructed skid-trails, to 550 for skid-trail segments approaching log landings.
The computer model successfully generated an optimized skid-trail network (Fig. 2b) connecting all but six log-piles to the three log-landings. These six log-piles were in areas with terrain slope around 75%, which was above gradient allowed for feasible skid-trails. However, as done with the operator-de signed skid-trail network, these log-piles were con nected to the closest optimized skid-trails and their associated skidding costs were also calculated. The number of loaded machine passes ranged from one to 619 indicating that more traffic was concentrated along fewer skid-trails arriving at the landings. Total skidding cost for the computer-generated network was slightly lower than the operator-designed net work ($67,563 vs $69,520, Table 1). The average skidding cost and average skidding distance per log pile was also slightly lower for the computer-gener ated network. However, skid-trail construction cost for the computer-generated network was higher ($13,447 vs $8,178) than the operator-designed net work. This was because numerous skid-trail seg ments were located across steeper terrain slope, which increased construction costs. On average in the operator-designed network, skid-trails with fewer than 20 loaded machine passes were placed on areas with terrain slopes of about 29% and skid-trails with more than 20 machines passes were located on areas with terrain slope of about 10%. On the other hand, same traffic level skid-trails, in the computer-gen erated network, were located on areas with terrain slopes of 43% and 19%. Thus, the resulting total harvesting cost for the operator-designed network was lower than the computer-generated network, 5 Comparing operator-designed and computer-generated skid-trail networks  Comparing operator-designed and computer-generated skid-trail networks approximately $77,700 and $81,000 or 4.8 $ ton -1 and 5.1 $ ton -1 . The total length of skid-trails in the computer generated network was 53.7 km, which is 2.0 km higher than the length of the operator-designed net work. This is likely because log-piles in the operator designed network were connected directly to the constructed skid-trails without terrain slope con straints and feasible skid-trails in the computer-gen erated network were allowed only when gradient was below 45%. As most of the skidding costs will be accrued while travelling along the high-traffic skid ding routes, the correct location of these paths is crucial because of their large impact on total costs. While the computer-generated results provide infor mation for the entire skid-trail network, the ground implementation of correctly identifying skid-trails connecting individual log-pile locations would be relatively difficult. A more practical application of the computerized model can focus on high-traffic or major skid-trails, which can be flagged on the ground to guide operator before construction. In this context, when comparing the length of skid-trails with more than 20 loaded machine passes, the computer-gener ated skid-trail network was about 28% shorter (9.4 km vs 13.1 km). This indicates that the computer generated network has a lower density of high-traffic skid-trails throughout the harvest unit concentrating skidding along fewer skid-trails. This becomes evident when comparing major skid-trails, those with more than 20 loaded machine passes ( Fig.2c and 2d). For example, the operator-design network has several major skid-trails arriving at the southern and northern landing following a parallel pattern, while the com puter-generated presents fewer major skid-trails fol lowing a branching pattern.
Several of these major skid-trails in the comput er-generated network also follow ridge lines and branch downslope to follow routes along contour lines. The branching pattern of the major skid-trails and of the entire skid-trail network is typical of studies using a network approach to determine routes that minimize skidding and construction costs (Stückelberger, 2008;Ezzati et al., 2015). The computerized model should also incorporate a two dimensional smoothing routine to help ensure a path that a loaded skidder can efficiently navigate. This would also reduce skid-trail length, which would also reduce skidding and skid-trails construction cost.
Lastly, there was a dramatic difference in procure ment area and volume received among log-landings. About 60% of the total volume was skidded to the northern log-landing, only 15% skidded to the middle log-landing, and the remaining 35% to the southern log-landing (Fig. 2b). The uneven volume distribution among log-landings and the relatively long average and maximum skidding distances (Table 1) suggest that overall skidding productivity and cost could have been improved by relocating both northern and middle log-landings farther north to reduce skidding dis tances.