A cost-benefit analysis of pre-sorting using a feller-buncher in underdeveloped short rotation poplar plantations

ABSTRACT Underdeveloped tree crops (⩽ 30 bone-dry tons ha−1) offer a main harvest of biomass-trees, which lack the size and the form for producing a log. However, about 1/3 of the available stems may yield at least one log, which could significantly increase the overall value of the harvest. Under those circumstances, it may make sense to sort log-trees from biomass-trees at an early stage, rather than having a harvester or a processor go through all of the treesonly to toss almost 70 out of 100 trees into the biomass pile. The Authors set up a controlled experiment to quantify the eventual benefit obtained by early pre-sorting, performed by the feller-buncher. Pre-sorting resulted in a 15% productivity loss for the feller-buncher, which was repaid by a 100% productivity increase for the processor. Considering the different hourly cost of the two machines, three Euros were saved for each additional Euro invested in pre-sorting. Pre-sorting made whole-tree harvesting (WTH) a significantly cheaper option than cut-to-length (CTL) , whereas the two systems would be almost equally expensive when no pre-sorting was applied. Pre-sorting would also facilitate multiple-tree delimbing, implemented through the introduction of a chainflail; that is likely to further reduce harvesting cost, returning financial viability to the harvesting of underdeveloped plantations.


Introduction
Establishing and managing a tree farm is a costly venture that accrues reasonable profit only if value recovery is maximized. That is the rationale behind traditional poplar plantations, which base their profitability on a rich harvest of valuable veneer logs and are still quite successful worldwide (Heilman 1999). In contrast, most tree farm models designed to produce low-value biomass have failed, unless additional revenues could be accrued from other services, such as phytoremediation, recreation, environmental protection, etc. (Helby et al. 2006). For that reason, most of the new plantations have been designed for the production of higher-value wood sorts than just industrial biomass (Stanton et al. 2002;Werner et al. 2012). However, not all plantations are as successful as their planners hoped: normally, a certain proportion of the total acreage established by any one company will grow less than expected and at the end of its rotation will not yield the quantity nor the quality than one hoped for. That is even more frequent when the plantation model is new and/or it is introduced to a new region, because of the uncertainty implied by the untested association between species and sites. In those instances, a common solution is whole-tree chipping, which allows minimizing cost (Spinelli et al. 2013). That generally achieves cost-cutting, not profitabil-ity or even break-even. On the other hand, trying to recover the small amount of higher-value assortments available amidst a score of biomass trees may incur a marginal cost higher than the marginal revenue (Pasanen et al. 2014;Karttunen et al. 2016). Picking out value trees from a mass of low-quality individuals is time consuming, and it can be too expensive -especially if the machine tasked with that job is a costly one (Kärhä 2011). For that reason, quality sorting should be done with the least expensive piece of equipment in the fleet, if at all. Furthermore, favorable conditions should be sought, in which the equipment tasked with quality sorting may operate the best. Past studies suggest that such could be the case of pre-sorting during felling (Lahrsen et al. 2021). That is the stage in the harvesting chain where the operator has the best view of the tree being cut and can easily assess its potential. Furthermore, the tree must be cut anyway, so that the eventual sorting task would simply amount to dumping cut trees into separate piles depending on whether they can offer some valuable assortment or just raw biomass. Performed at any other stage along the chain, sorting would require breaking the tree out of a bunch or stack, evaluating whether it is good or not and eventually cutting off a log from those few trees that turn out to have the right form and size. Therefore, pre-sorting during felling may seem a promising technique for value recovery in those plantations that did not grow as hoped for and may offer too meager a harvest for standard collection, whether by the cut-to-length (CTL) or the whole-tree harvesting (WTH) system. The goal of this study was to perform a cost-benefit analysis of pre-sorting during felling, conducted in a poplar plantation representative of the new tree farms being established in Central and Eastern Europe (IPP 2019; Heilig et al. 2021). In particular, the study compared conventional cut-to-length and whole-tree harvesting without pre-sorting against whole-tree harvesting with presorting, conducted by the same feller-buncher tasked with tree felling. The treatments on test were then: 1 -CTL: Conventional cut-to-length harvesting, without pre-sorting (single-tree) 2 -WTH: Conventional whole-tree harvesting, without presorting (multi-tree) 3 -WTH Pre-sorting: Whole-tree harvesting with presorting (multi-tree) Pre-sorting was not tested in association with cut-to-length harvesting because that would require introducing an additional machine to the CTL system and likely cancel any benefits eventually accrued by pre-sorting. Pre-sorting was only attempted where it could be done by introducing a small procedural change to an established routine, for fear that any stronger interventions could irremediably compromise the already precarious balance. In all cases, the target assortment was 4-m long logs with a minimum small end diameter (SED) of 7 cm. The ultimate goal was to produce the largest proportion of logs at the lowest possible cost. Therefore, the comparison was based on the following key performance indicators (KPI): • harvesting cost • log recovery rate.
The null hypothesis was that there were no statistically significant differences between the three treatments on test for each of the two KPIs.

Materials and methods
The three treatments were tested in January and February 2022 on one of the plantations established in Western Slovakia by IKEA Industry for supplying their particle board factory in Malacky. The test plantation was located in Pernek (48° 23' 55.04" N; 17° 04' 43.97" E in WGS84), a few kilometers northeast of the plant (48° 25' 01.81" N; 17° 02' 06.95"E). It was a 6-year-old poplar stand established at a rectangular spacing of 3.0 m × 2.0 m with hybrid poplar (Populus x euramericana Dode (Guinier)), using the following mix of similar clones: AF13, AF 16, and AF18 (Landgraf et al. 2020;Meyer et al. 2021). The trees were planted on an ex-arable field, on flat and even terrain. Weather during the test was consistently warm and dry, with occasional light rainfall. The air temperature varied between −2 and +14 C°. All operations were conducted in daylight, normally from 07:00 to 17:00.
Under the WTH treatment, trees were felled and bunched using a 20-t Kobelco GS200 tracked excavator equipped with a Silvaro 250 K double-blade shear. Whole-tree bunches were taken to the roadside landing using a Ponsse Buffalo heavy forwarder (140 kW, 15 t own weight, 12 t payload capacity https://www.ponsse.com/products/forwarders/product/-/p/ buffalo_8w#/). That machine was large enough to load wholetree bunches into its basket and carry them off the ground, to the benefit of minimal soil contamination. Once there, trees were picked off the piles and processed into 4 m lengths using a Nisula 425 C processor fitted to a Volvo EW 160 wheeled excavator. The landing was obtained at the field's edge, along a service road, and was large enough for the processor to operate comfortably. Logs and tops were piled in separate stacks along the road, for later loading (logs) or chipping (tops) into trucks.
The pre-sorting treatment was essentially the same as the WTH treatment, except that the feller-buncher operator would pile log-trees and biomass-trees into separate piles, which would be extracted separately with the Ponsse forwarder and unloaded into separate piles. Hence, the processor would only work on the pre-sorted tree piles and spend less time picking up and casting out biomass-trees. Piles consisting of biomassonly trees would be chipped later, together with the piles of tops, branches, and undersize trees discarded by the processor.
Each machine was driven by its own operator, which was a qualified professional with significant experience of his machine and specific task. All operators were co-opted into the study after observing over a dozen contractors and casting out those that were not considered representative of the local pool of professional machine operators (Spinelli et al. 2022a).
Before engaging with data collection, all operators were tasked with working at least one full day in the plantation outside the marked experimental blocks. That way, they could adjust to the study routine and methods. The harvester and processor head measuring systems were checked and calibrated on 10 trees before starting the experiment. Afterward, calibration was checked on 2 trees at the beginning of each working day (Spinelli et al. 2022a).
The experimental design was a factorial scheme where each treatment was repeated 8 times ( Figure 1). Each repetition consisted of one sample plot measuring approximately 15 × 50 m (700 m 2 ). Plot width was selected to represent the work frontage of the harvester and the feller-buncher, which cut 5 row bands (5 × 3 m = 15 m). Plot length was chosen for including approximately 130 trees, which previous studies of the same machine types found to be the amount of work normally performed in about one scheduled machine hour (SMH). The experimental plots were alternately assigned to the three treatments, on the assumption that a pure randomized design could confuse the operators and increase the potential for attribution errors. In contrast, the alternate checker-board design still allowed for an even spread of eventual site gradients but held less potential for confusion (Spinelli et al. 2022a). The beginning and the end of each experimental plot were clearly marked with bright paint to facilitate attribution. The circumference at breast height of all trees in all plots was measured manually with a measuring tape and then converted into diameter at breast height (DBH), over bark. Furthermore, 6 trees covering the whole DBH distribution were destructively sampled to determine total height and weight, separately for the assumed log and chip portions (Urban et al. 2015;Krejza et al. 2017). Destructive sampling allowed estimating an allometric equation representing the relationship between DBH, total height, and mass: that equation was then used to predict the standing mass available on each individual sample plot (Headlee and Zalesny 2019). Previous research has demonstrated that it is possible to build reliable allometric models with such a small sample, when tree variability is as small as normally found in even-aged clonal plantations (Hartmann 2010;Verlinden et al. 2013;Hjelm 2015). In any case, mass estimates were later adjusted using adhoc correction factors that were obtained by matching the preharvest inventory data with the actual harvest taken to the weighbridge available at the receiving plant in Malacky. That was done separately for the log and for the chip portion obtained from each of the three treatments, in order to account for variations in log recovery that might be associated with any specific treatments (6 correction factors). Water mass fraction was determined both at the time of destructive sampling during the pre-harvest inventory and at the time of scaling on the factory weighbridge, so as to match dry mass pre-harvest estimates with dry mass post-harvest weighbridge figures. In both cases, water mass fraction was determined with the gravimetric method, according to EN ISO 18134-2:2015. Mean water mass fraction at delivery was 55% (standard deviation = 2.7%).
Depending on site and treatment, the ratio between factory dry mass and inventory dry mass varied from 0.63 to 0.86 with an overall average at 0.66 -meaning that the field inventory overestimated actual harvest by about 50%.
At the time of harvesting, researchers determined for each sample plot the time taken to treat each individual plot by each individual machine, using a stopwatch accurate to the second. Both productive and delay times were recorded (Björheden et al. 1995), but delay time was eventually excluded from the study and replaced by a 20% delay factor. That was done because the time spent on each sample plot was not long enough to accurately estimate the long-term incidence of delay time. The 20% increase applied to the data was consistent with the findings of previous published studies, with special reference to the harvesting of plantation forests (Spinelli and Visser 2008). That figure was also quite close to the sum of all delays recorded during the complete study, as conducted on the 24 experimental plots.
Since the capacity of both forwarders would often exceed the mass of wood on one experimental plot, a payload fill rate was estimated for each trip and applied as a reduction factor to travel time, on the assumption that a forwarder would normally travel with a full load. The inherently subjective character of that estimate was minimized by taking pictures of each load and submitting them to a panel of five experts, together with a reference picture of the forwarder with a full load. The experts were then asked to give their own best estimate of the payload fill rate for each load, and the value eventually adopted into the study was the one receiving most preferences, or the central one if each expert returned a different figure.
All wood harvested from the experimental plots was taken to the same landing, and therefore the extraction distance was the same for all treatments evaluated. In any case, distance was estimated from Google Earth pictures, from the center of the sample plot to the center of the roadside landing and it was 370 m as average. Distance between the strip roads was 15 m. Mean load concentration was 3.3 BDT (or 7.3 green t) per 100 m. The plot-level time study was accompanied by a parallel cycle-level elemental time study (Magagnotti et al. 2011) that covered about half of the harvester, feller-buncher and processor cycles on each sample plot, plus all the forwarder cycles (Annex A). The goal of the elemental time study was to determine if treatment would specifically impact one or more work steps within the complete task; in particular, it aimed to determine whether and how pre-sorting would affect the work pattern of the feller-buncher and the processor, which were the units directly impacted by it. Similarly, the cycle-level time study was used to determine if pre-sorting would produce different tree rejection rates (i.e. larger or smaller proportions of no-log trees) and thus explain possible differences in log yield.
Machine cost was assumed to be the rates actually charged by the service providers. These were: 69 € per scheduled machine hour (SMH) for the harvester, 53 € SMH −1 for the compact forwarder, 45 € SMH −1 for the feller-buncher, 75 € SMH −1 for the heavy forwarder and 47 € SMH −1 for the processor. The Authors acknowledge that the commercial rates charged by individual operators can hardly offer a general reference and they encourage readers operating under substantially different economic conditions to recalculate harvesting cost using their own rates and the productivity data presented in this paper. Otherwise, one could use official rates, but those were not available for the region where this study was conducted. In any case, the rates charged by the operators contracted for this study were in line with those charged by other service providers in the region, and therefore the selected rates can be taken as generally representative of the area (Spinelli et al. 2022a). However, those rates were much lower than those charged by neighboring Austrian contractors, located just across the border -which stresses the importance of recalculating cost whenever market conditions differ significantly from those described in this paper (Spinelli et al. 2022b).
Data analysis aimed at estimating meaningful measures of centrality for each of the three treatments tested, and at determining if the differences found between them were statistically significant. The dataset comprised 8 repetitions per treatment, each repetition represented by one sample plot. Given such relatively small number of data and the frequent violation of the normality assumptions, the analysis was conducted with non-parametric techniques, which are robust against such violations and still reasonably accurate. In particular, the Kruskall-Wallis multiple comparison test was used for a general comparison between treatments, while the Mann-Whitney unpaired comparison test was used for comparing the unsorted against the pre-sorted WTH treatment, and the least expensive between the two WTH treatments against the CTL treatment. For all analyses, the significance level was set at α < 0.05. The analyses were implemented with the software SAS Statview. In order to facilitate comparison between the WTH and CTL treatments, the separate time consumption and cost figures obtained for the feller-buncher and the processor under the WTH treatment were consolidated into one single felling and processing time consumption and cost figure per sample plot, which was functionally equivalent to the felling and processing work sequence conducted within a single pass by the harvester, under the CTL treatment. That way, one would have one figure for felling and processing and one figure for forwarding for each of the sample plots, regardless of treatment.

Results
Median diameter at breast height (DBH) was 10 cm and median height 8 m, without any significant differences between treatments. No significant differences were found for tree mass and stand stocking either, with medians at 14 kg dry matter (DM) and 22 bone dry tons (BDT) per hectare, respectively (Table 1). Log yield was low, with medians between 17% and 25% of the total standing mass; it was higher for the CTL treatment, but that difference was not statistically significant and the information can be considered as suggestive rather than conclusive.
Cumulated felling and processing productivity was significantly higher for the CTL harvester, while cost was significantly lower for the WTH pre-sort treatment. The latter accrued a 20% cost saving, compared with both alternatives (Table 2). Forwarding was significantly more productive under the WTH treatment, since a larger and more powerful machine was deployed. However, that machine was also more expensive and the productivity margin (12%) was smaller than the additional cost per hour (40%), resulting into a significantly higher forwarding cost for both WTH treatments. That eroded the benefits accrued during felling and processing and resulted in a tied match when it came to the overall harvesting cost (i.e. felling, processing, and forwarding: from standing trees to logs and biomass at the roadside landing). In fact, the data suggested a harvesting cost stratification whereby WTH > CTL > WTH Pre-sort, but those differences were not significant at the 5% level.  In fact, forwarding was somewhat inefficient under all treatments, given the difficulty of assembling large enough loads with the lightweight bulky materials available on site. Mean full payload was 2.8 BDT, 2.2 BDT and 4.0 BDT, respectively, for the Logset forwarder (CTL) loaded with logs and with tops, and for the Ponsse forwarder (WTH). Expressed as fresh weight, those figures would amount to 6.2, 4.8, and 8.8 green tons, respectively. That is: 55%, 40%, and 60% of the maximum rated payload of the respective forwarder. Even with 4-m logs under the CTL treatment, the compact forwarder was unable to pack more than 60% of its rated payload -and logs represented only 25% of the total mass on site, the rest being much bulkier tops and branches! The specific benefits of pre-sorting are best appreciated through a direct comparison between the two WTH treatments, conducted exactly under the same conditions: site, machines, and operators (Table 3). Essentially, pre-sorting resulted in a 15% productivity loss for the feller-buncher, which was repaid by a 100% productivity increase for the processor. The additional felling cost was 2.7 € BDT −1 , but that allowed saving 8.8 € BDT −1 in processing cost; three Euros were saved for each additional Euro invested in pre-sorting, which seemed a pretty good deal. Those differences were statistically significant, whereas the difference in forwarding cost was not. However, that difference went in the opposite direction than that found for felling and processing, which increased the variability of the overall cost figures to the point where the statistical analysis turned inconclusive. Therefore, the 15% saving on overall harvesting cost accrued through pre-sorting could only be taken as suggestive (p = 0.14). However, when the data were corrected by replacing the actual forwarding cost with its overall average across treatments, then the analysis confirmed the statistical significance (p = 0.03) of that same cost saving figure (15%).
The detailed cycle-level time study was unable to detect a significant productivity difference for the feller buncher under either of the two treatments (Table 4). Under the unsorted treatment, larger accumulations (3 trees per cycle instead of 2) were significantly more frequent than under the pre-sorted treatment. However, that incurred a proportionally larger time consumption, so the productivity was approximately the same. Taken together, the results of the sample plot-level and the cycle-level studies indicated that any productivity losses caused by the pre-sorting task were small or altogether negligible.
The detailed cycle-level time study showed that processor productivity in trees per hour was slightly higher (13%) under the unsorted treatment, compared with pre-sorted one, and that the difference was statistically significant (Table 5). That was the logical consequence of more trees being picked up, checked, and cast away into the biomass pile without processing, which resulted in the average time consumption per tree being shorter under the unsorted treatment. What really made the difference was that after pre-sorting 60% of the trees in the processor piles were logworthy, while that number dropped to 30% if no pre-sorting was performed. In short, the original crop consisted of 70% biomass-trees, all of which had to be picked up by the processor just to be chucked into the biomass pile, thus leading to an accumulation of void cycles that would not produce any logs and represented wasted processor time. Pre-sorting resulted in the processor having to handle 20 biomass-trees for every 30 log-trees, that is: 50 cycles (30 log-trees + 20 biomass-trees) instead of 100 (30 log-trees + 70 biomass-trees) under the unsorted treatment. It was such Obs. = 8 replications per treatment -sample plot-level study; "Forwarding corrected and equalized" means that the same mean forwarding cost was adopted for WTH regardless of pre-sorting, since there was a difference in forwarding cost between WTH unsorted and pre-sorted, but that was not statistically significant.  a dramatic reduction in the number of processor cycles required to treat a given lot that accrued the large savings shown in Table 3. The direct match between CTL and WTH harvesting (presorting) indicated the latter as the likely winner, with an 8.5% cost saving (i.e. 4 € BDT −1 - Table 6). That difference was borderline significant (p = 0.06), but given the relatively small number of observations it could be taken as strongly indicative of a real stratification of the data, which could also be observed in the box plot ( Figure 2).
Essentially, switching from CTL to WTH harvesting (presorted) accrued a 7.5 € BDT −1 saving in felling and processing cost and incurred a 3.6 € BDT −1 increase in forwarding cost. Savings were twice as large as costs, which should logically result in an overall reduction of the harvesting bill: that was one more reason why one should consider with much attention the borderline significant difference reported for total cost in the last row of Table 6.

Discussion
Before getting too deep into the discussion it is fair to acknowledge the main limitations of this study, so that readers can decide to what extent they want to trust our conclusions as a guide for future practical use. We believe that a main limitation of this study is the small number of replications, which may have been too few for emerging the finer differences between treatments. However, while the observations were relatively few, each of them contained enough work for achieving stable performance levels. Therefore, the experiment was able to disclose significant differences and offer good insights into the selected systems. Furthermore, the raw data were not manipulated in any way, in order to keep their genuine value and avoid introducing any bias. For that reason, nonparametric statistics were preferred to transformations. The idea was that readers should be able to follow and understand the whole process throughout, without complicated analyses or mysterious black boxes.
A further limitation of the study was the inevitable impact of operator effect (Purfürst and Erler 2011;Leonello et al. 2012). Ideally, that should have been included into the experiment by recruiting multiple operators (Spinelli and Moura de Arruda 2019), but such intervention would have dramatically expanded the cost and the complexity of the experiment. For that very reason, the inclusion of multiple operators is a very rare occurrence in forest operations studies -whether controlled or observational. In that regard, recommended best practice consists of selecting competent professionals that are representative of the larger pool of available operators, after observing them at work (Košir et al. 2015). That was done here; even if there is no guarantee that the results of the system comparison (CTL vs. WTH) could not have been confounded to some extent by minor variations in operator competence, we believe that operator effect was unlikely to be as strong as to overpower important performance differences. Furthermore, operator effect may have confounded the comparison between CTL and WTH, but not that between unsorted and sorted treatment, given that the very same operators were deployed for both treatments. Support to cautious generalization is obtained from the existing literature. Few references can be found that specifically address the effect of pre-sorting on feller-buncher performance, but all those few agree with the direction and the size of effects reported here; they indicate that pre-sorting leads to a small but significant reduction of feller-buncher productivity, estimated between 5% (Mikkonen 1976;Gingras 1996) and 10% (Gingras and Godin 2001;Kizha and Han 2016). In fact, differences may even be non-significant when only two sorts are separated (McMorland 2008). The literature also indicates that pre-sorting by a feller-buncher significantly increases the productivity of a processor (Meek 1995), which would incur larger losses than the feller buncher, if forced to sort its feedstock (i.e. 15-30% losses according to Gingras 1996).
In fact, the crux of the matter is the small proportion of log-trees offered by underdeveloped plantations. Separating felling from processing and delegating the two tasks to separate machines is generally done in the hope that dedicating a specialized and cheaper machine to each individual task will result in a higher productivity and lower cost, so that the sum of the two separate costs will be lower than the cost incurred when deploying an integrated single-pass cutto-length harvester for the same task. That is always a tough match between the principle of specialization and that of redundancy -since the two separate machines will result in the same work object being picked up and laid down twice, which is not inherently efficient. In our case, however, task  separation resulted in redundancy only for a small proportion of the trees, since pre-sorting allowed setting aside those trees that did not need a second pass by the processor. So, task separation did not cause redundancy, but rather avoided it -because it prevented the redundant delimbing and measuring performed by the CTL harvester, which would end with rejection 70% of the times.
In that regard, we cannot fail to notice that even after presorting, 40% of the trees fed to the processor were going to be cast away, all the same. While that was a large improvement over the unsorted treatment, one still wonders if the fellerbuncher operator could make a better selection without sacrificing too many log trees. Further research could address the use of sensors for assisting with tree size estimate and enable a more accurate selection: that way one could reduce the proportion of biomass-trees ending into the log-tree pile, without increasing the wastage of valuable log-trees (Forsman et al. 2016).
In fact, log recovery must also depend on harvester or processor head characteristics, as well as on the individual skill of the operators. We have already explained why we believe that the operators selected had similar good skills and were representative of the larger operator population available in the region; however, the two machines were definitely different. The dedicated harvester was more technologically advanced (and expensive) than the excavator-based processor. Visual observation during the extended time study suggested that the processor head design and setting were less tolerant of malformed trees, which struggled to pass through the knives and would often break. Therefore, one may wonder if significant improvements in log yield and productivity could be obtained by replacing the currently used processor head with another model, possibly more suited to the task (Spinelli et al. 2002;Magagnotti et al. 2021). In that regard, a more radical solution would consist of replacing the processor with a multitree delimber, such as a chainflail. The latter would work best with pre-sorted small trees, likely offering higher productivity and log yield than the single-tree equipment used in this test. Use of a chainflail may allow bumping many borderline trees into the log-tree pile, since that machine is more tolerant of stem malformation than most cut-to-length heads (Spinelli et al. 2022c). In that regard, it is worth mentioning that the market offers multi-tree CTL harvesters and processors that could be used profitably for mass-processing pre-sorted trees or conduct pre-sorting and mass-processing at the felling stage. They were not included in the test because those machines are still relatively few (none was available in the area) and because previous tests on short-rotation poplar indicated that the production benefit they can accrue is relatively small (Magagnotti et al. 2021). Finally, one may also try to cut the cost of CTL harvesting by replacing the thinning processor with a light excavator-based or tractor-based harvester. However, such equipment is often less productive than its dedicated counterpart and it is generally resorted to when market conditions prevent full utilization of a specialized machine so that investment cost must be minimized (Bergroth et al. 2006).
Further improvements may target forwarding. The study highlighted three important issues when it came to forwarding: first, the very low payload utilization achieved for all treatments evaluated; second, the better performance achieved by the lighter of the two machines; third, the possibly better performance achieved by the heavier machine under the unsorted treatment, compared with the pre-sorting treatment. First things first: low payload utilization is a common challenge when forwarding residues and whole trees, due to their bulky nature; however, here that happened with log loads, too. IKEA grows poplar for manufacturing an innovative lightweight board -so, low density is the essential characteristic of the wood obtained from those plantations: there is no way around it. Forwarders used in short rotation poplar (SRP) do not need to be particularly powerful or robust: they will inevitably carry relatively light loads and will travel on easy terrain, with no important obstacles or inclines that may tax the smallest engine. If one could expand the load space of a thinning forwarder, that would be more than enough, and it may even be worthwhile designing a dedicated SRP model -lighter and cheaper than most standard units. The potential benefit of such an approach is confirmed by the second issue in the list: the heavy forwarder incurred a higher extraction cost than the smaller compact forwarder, because the advantage offered by its larger size and power could not be brought to bear and the higher hourly rate was not offset by a proportionally higher productivity. Finally, one should consider the better performance obtained by the same heavy forwarder under the unsorted WTH treatment, compared with the pre-sorted WTH treatment. There, data were not conclusive, but only suggestive. Yet, they pointed at something that has already been reported in other studies: the inverse relationship between forwarder productivity and biomass stocking (Väätäinen et al. 2006;Manner et al. 2013). As applied in this experiment, pre-sorting effectively resulted in a reduction of biomass stocking, since the forwarder had to make two passes for the separate collection of the two tree sorts. There, redundancy could be avoided by introducing a partitioned load bay, offered as optional by several manufacturers. Such modification would allow packing multiple sorts into the same load in a single pass and speed up extraction. Essentially, SRP requires a small forwarder with a very large bunk, fitted with a mobile partition; such machine would allow rebalancing cost with performance, thus maximizing costefficiency. Since whole trees must be handled, the loader should be quite powerful and it could also be fitted with a heel, to improve control of long loads (US Forest Service 2022). Introducing a heel would incur a small additional cost but could improve handling efficiency and reduce breakage risk, especially when dealing with the brittle stems offered by some low-density clones (Zhang et al. 2003).

Conclusions
Pre-sorting of log-trees offers significant efficiency benefits when harvesting underdeveloped short rotation poplar. First and foremost, it avoids burdening the cut-to-length processor with fruitless work that does not result in the production of any logs. Furthermore, it facilitates the introduction of multi-tree delimbing technologies, which could result in a marked increase of productivity and log yield. In that case, WTH would likely outperform CTL harvesting and help take harvesting cost back within the profitability limits set by the operation managers. Of course, one should always balance the additional cost of integrated harvesting (logs + chips) against the savings accrued through wholetree chipping, after accounting for the amount of log material actually recovered and the price difference between logs and whole-tree chips.
Annex A -Description of the time elements for the cycle-level study A1. Harvester Move = Wheels are turning while no other tasks are being performed Fell = From the moment the boom extends toward a tree to the moment when the stem starts being fed through the head Process = From the moment the stem starts being fed through the head to the moment the last log is cut Drop top = From the moment the last log is cut to the moment the top is released onto the ground or pile Stack = Logs and tops are moved and/or rearranged A2. Processor As for the harvester, except that Fell is replaced by Grab = From the moment the boom extends toward the tree pile to the moment when the stem starts being fed through the head A3. Feller-Buncher Move = Tracks are turning while no other tasks are being performed Fell = From the moment the boom extends toward a tree to the moment when the tree or tree bunch is moved toward the dumping site Handle = From the moment the tree or tree bunch is moved toward the dumping site to the moment the head is empty and is positioned toward the work frontage.
A4. Forwarder Travel empty = The forwarder leaves the landing and moves to its first loading station Loading = From the moment the forwarder reaches its first loading station to the moment when it has completed its load and moves toward the landing Travel loaded = The forwarder leaves the loading site and reaches the roadside landing, stopping by the designated stack Unloading = From the moment the forwarder stops by the designated stack to when it leaves the landing