A procedure for automated tree pruning suggestion using LiDAR scans of fruit trees
11 A procedure for automated tree pruningsuggestion using LiDAR scans of fruit trees.
Fred Westling*, James Underwood, Mitch Bryson (cid:70)
Abstract —In fruit tree growth, pruning is an important managementpractice for preventing overcrowding, improving canopy access to lightand promoting regrowth. In fruit with a high energy content, includingavocado (Persea americana), ensuring all parts of the canopy havesufficient exposure to light is of particular importance. Due to theslow nature of agriculture and the numerous parameters contributingto yield, decisions in pruning, particularly in selective limb removal,are typically made using tradition or rules of thumb rather than data-driven analysis. Many existing algorithmic, simulation-based approachesrely on high-fidelity digital captures or purely computer-generated fruittrees, and are unable to provide specific results on an orchard scale.We present a framework for suggesting pruning strategies on LiDAR-scanned commercial fruit trees using a scoring function with a focus onimproving light distribution throughout the canopy. Due to the destructivenature of physical experimentation, this framework is presented usinga three-stage approach where stages can be independently validated.Firstly, a scoring function to assess the quality of the tree shapebased on its light availability and size was developed for comparativeanalysis between trees using observations from agricultural literature,and was validated against yield characteristics from an avocado andmango orchard. This demonstrated a reasonable correlation against fruitcount, with an R score of 0.615 for avocado and 0.506 for mango.The second stage was to implement a tool for simulating pruning byalgorithmically estimating which parts of a tree point cloud would beremoved given specific cut points using structural analysis of the tree.This was validated experimentally using manually generated groundtruth pruned tree models, showing good results with an average F1score of 0.78 across 144 experiments. Finally, new pruning locationswere suggested by discovering points in the tree which negativelyimpact the light distribution, and we used the previous two stages toestimate the improvement of the tree given these suggestions. Theseresults were compared to a tree which was commercially pruned usingexisting wisdom. The light distribution was improved by up to 25.15%,demonstrating a 16% improvement over the commercial pruning, andcertain cut points were discovered which improved light distribution witha smaller negative impact on tree volume. The final results suggest valuein the framework as a decision making tool for commercial growers, oras a starting point for automated pruning since the entire process can beperformed with little human intervention. Further development should beperformed to improve the suggestion mechanism and incorporate moreagricultural objectives and operations. Keywords —agriculture; lidar; pruning; orchard; suggester
All authors are with the University of Sydney*Correspondence: [email protected]
NTRODUCTION
Tree pruning is an important part of fruit tree orchardmanagement, but must be done carefully. Miller et al.(1960) demonstrated that severe pruning can reduce theyield of the current crop, but increase it in followingseasons. As such, it can be a difficult process to informand feed back for improved decisions. Since tree growthtakes many years, there is interest in performing treeanalysis in silico to judge decision making or allow studyof physically challenging or destructive operations. Inthis work we focus primarily on avocado and mangotrees, though methods developed are based on researchfrom a variety of fruit, and should be applicable to anyfruit for which light availability is important in growth.Pruning can be undertaken for a variety of reasons,including lowering tree height to reduce harvesting costsand reduce mutual shading of scaffold branches (Thorpand Stowell (2001)) as well as reducing inter-tree crowd-ing, promoting regrowth and risk mitigation (Purcell(2015)). Partida Jr (1996) observed that pruning ’Hass’avocado trees produced healthier trees and grew largersized fruit, and that maintaining the trees at a lowerheight increased overall tree vitality and fruit productioncompared to trees which were allowed to grow at will.Beyond individual tree health, pruning also improvesproductivity by reducing crowding. Wilkie et al. (2019)showed that, in conventional large-tree low-density or-chards, the yield/ha increases with total light intercep-tion and canopy volume up to a certain point, and thendeclines. In avocado trees in particular, the two primarymethods of pruning are ”hedging” (planar cuts) and”selective limb removal”. Menzel and Le Lagadec (2014)noted that hedging is most effective in warm subtropicalcoastal areas, while limb removal is more effective in cooltemperate areas where successive crops overlap. Figure 1shows an example of conventional avocado trees whichhave undergone both processes, significantly reducingtheir size and quantity of leaf matter to stimulate futureyields.Given the slow nature of growing studies and theinterdependent variables contributing to yield, pruningrecommendations are typically drawn from old practicesor studies in similar tree crops. Our focus is on pruningdecisions designed to improve light penetration through a r X i v : . [ c s . C V ] F e b (a) Unpruned, camera (b) Pruned, camera(c) Unpruned, labelled LiDAR (d) Pruned, labelled LiDAR Fig. 1: Example of pruning applied to an avocado tree.The LiDAR scans were captured at the same time thephotos were, and were then manually labelled with trunkand foliage and raytraced with simulated light for clarity.the canopy, which Stassen et al. (1999) lists as a keyconsideration in fruit production. For avocado trees inparticular, Mickelbart (2004)notes that the leaves respondrelatively slowly to light, so increasing access to sunlightthrough pruning should improve fruit quality and yield.Snijder and Stassen (1995) showed that light penetrationwithin the canopy can be improved from 7% to 58%through selective pruning, while Stassen et al. (1999)notes that maximum photosynthesis occurs at 30% ormore of full sunlight intensity. This was further validatedby the findings of Marini et al. (2020), who showedthat peach and other high-sugar fruit require all areasof the tree to achieve 25% of full sunlight for optimumgrowth. The decisions used to open up the canopy tendto be ad-hoc, though care must be taken to balance otherpruning considerations. For instance, ¨Orn (2016) founda correlation between total light and fruit quality, butalso a strong link between volume and yield. Thorp andStowell (2001) demonstrated that tree height could bereduced to 6m without negatively impacting yield, butnot to 4m. In general, fruit tree pruning is a process ofidentifying specific goals and shaping the tree to achievethem while minimising negative effects.Functional-structural plant modelling (White andHanan (2012, 2016)) has been used to generate virtualtrees using algorithmic growth to study parameters in-cluding light, nutrients and spacing. Da Silva et al. (2014)and Tang et al. (2015) generate computer models of appleand peach trees respectively to study light interceptionefficiency, while Yang et al. (2016) used similar models toanalyse canopy structure. Others have used captures of real trees rather than purely virtual ones. Sinoquet et al.(2007) digitised trees in high quality using a magneticdevice to accurately recreate leaf angle distributions andplant components to correlate fruit-scale light intercep-tion with fruit sugar content. Methods like these tendto be difficult to perform at scale, while Westling et al.(2018) used fast hand-held laser scanners to generatelower-quality tree models which nevertheless allowedlight interception and distribution studies using publicweather data.Physical pruning is destructive by its very nature sovirtual approaches to analysing or predicting its effectscan be helpful, though few works have explored thisspace. By comparing point clouds at different times,Xiao et al. (2012) perform change detection of trees inurban areas using aerial scans. Similarly, Estornell et al.(2015) presented an approach to estimate the amount ofbiomass removed during pruning. Tagarakis et al. (2018)showed the potential to use LiDAR sensors for scanningolive trees prior to pruning to help inform manage-ment practices. Recently, Strnad et al. (2020) presenteda framework for optimising virtual tree pruning againstmultiple objectives by simulating the effect of pruningand the subsequent growth of new matter. However,their framework operates purely on virtual trees whichare fully observed.If tree pruning were to be automated, it must bepossible to make pruning decisions autonomously, inparticular the limb removal decisions which are spe-cific to individual trees that must be observed. Mobilescanning platforms like that presented by Underwoodet al. (2016) enable easy and automated 3D point cloudcapture of orchard trees, while the method of Westlinget al. (2018) allows rapid analysis of the light intercep-tion and distribution of a tree using low-quality pointclouds. Here we present a method for non-destructiveanalysis of the impact of pruning decisions, specificallyon the immediate light distribution characteristics on a3D point cloud model of individual trees as captured by ahandheld laser scanner. Due to the difficulty of capturingreal-world data during pruning and tracking commercialpruning decisions, we develop a virtual tree data setusing the SimTreeLS system presented by Westling et al.(2020a) which can be used to produce realistic LiDARdata upon which pruning analysis can be conducted,though we demonstrate the ability to transfer the methodto real data.
ETHOD
Real-world LiDAR point cloud data were capturedat Simpson’s avocado farm, a commercial orchard inChilders, Queensland. Avocado trees are well structuredfor pruning, with wide splits and long, curving branches.In 2017 and 2019, we captured data before all pruningand after pruning. In 2017 in particular we were able tocapture before hedging, before limb removal and afterall pruning, though we were not able to replicate this in
In order to compare tree quality, we developed a scoringfunction based on the literature presented in Section 1.Specifically, Marini et al. (2020) and Stassen et al. (1999)showed that ideal light distribution for fruit trees occurswhen all areas of the tree have access to 25-30% of avail-able sunlight, and ¨Orn (2016) found good correlationsbetween fruit yield, tree volume and total light absorbed.In order to balance these considerations, we propose ascoring function which promotes total light absorbed andtree volume while penalising leafy matter with accessto less than 25% of full sun. Figure 2 shows a pointcloud of a tree in which each voxel is coloured by thepercentage of light absorbed, with all blue and yellowvoxels demonstrating a sub-optimal access to light.First we compute the light environment of the tree byintegrating solar data across a specified growing seasonas described by Westling et al. (2018). The integrated lightenvironment is then raytraced through the voxelised treepoint cloud and the amount of energy absorbed by eachvoxel is recorded. The volume of the tree is calculatedby summing the convex hulls of connected componentsat different height levels, using a method presentedand validated by Westling et al. (2020b). The total lightabsorbed by the tree is computed by summing the lightabsorbed by each voxel. Given the tree is represented by N v voxels, each with a quantity of light absorbed L i , wethen calculate, for each voxel, the percentage of full lightabsorbed by that voxel: p i = L i max j =1: N v L j (1)We then calculate the ”distribution score” D of thetree in order to penalise voxels under 25%, following the Fig. 2: Point cloud of a tree, coloured by light distributionstatistic. Trunks points (perfectly known since this is asimulated tree) are coloured in black. Voxels with lessthan 25% of full sun absorbed are considered suboptimalfor production of high quality fruit.target distribution presented in Figure 3: D = 1 N v (cid:88) i =1: N v (cid:26) − (0 . − p i ) , if x ≤ . log ( p i + 1) , otherwise (2)Since our scoring function is intended for comparativerather than predictive purposes, we developed it to bea relative score and not an absolute one. With thatassumption, we normalise the volume and total lightabsorbed using the maximum value of each within thedata set in question. This prevents us from claiming ifa tree is ”good” or not in isolation, but allows us tojudge if one tree is better than another, or whether a treeimproves following a change. ˜ V = V i (cid:80) i =1: N trees V i ; ˜ L = L i (cid:80) i =1: N trees L i , (3)where V i is the volume of tree i and L i is the sum oflight absorbed by each voxel in tree i .Finally, we calculate the final score S for a given tree asthe linear combination of the three components, to allowtuning of the importance of each. S = αD + β ˜ V + γ ˜ L (4)We calculated the values of α , β and γ experimentally,using the reported yields of trees for which we have Fig. 3: Response function for mapping P value to D score.This function is designed to penalise values under 0.25and avoid overly rewarding voxels above 0.25. The redline denotes y=0point cloud scans. Fruit tree yield is obviously a complexvalue, dependent on a wide variety of factors. However,the results presented by ¨Orn (2016) and the insightsof Stassen et al. (1999) suggest it should correlate withan effective method for scoring the light distribution oftrees. For avocado trees, we had yield data for 18 datumtrees in one year. We also had yield data reporting froma mango orchard intensification trial on the WalkaminResearch Station, operated by the Queensland Govern-ment Department of Agriculture and Fisheries (DAF),comprised of 270 datum trees grown using differenttraining methods and densities, leading to a variety oftree shapes. There were 15 independent combinations ofindependent variables and 6 replications of each. Whenevaluating coefficients, we took the average yield andfruit weight for the 3 trees in each of these 90 experimentsto suppress outliers.Though our yield data represented different fruit treeswhich are different in size, shape and age, we found aset of coefficients with a similar performance for both,shown in Table 2.1.
Coefficient Value α β γ TABLE 1: Coefficients experimentally chosen for equa-tion4
In order to suggest cuts which improve the score of atree, we must simulate the effect of a particular prune on the structure of the tree. We did this using a methodpresented by Westling et al. (2020c) and illustrated inFigure 4. Each point cloud was voxelised to normalisematter density and speed up the processing, then a graphwas created taking the mean location of the points ineach voxel as the nodes, and creating edges between eachnode and its neighbours within a given radius. Given aspecified cut point, we marked the node at that pointand all nodes within a certain radius as part of the cut.The shortest path from each node to the trunk was thentraced using A*, and each node whose path includesthe cut nodes is taken as pruned, with this classificationpropagated to all points within that voxel.Fig. 4: Illustration of the pruning simulation method.An approximation of a generated graph is overlaid onthe point cloud, with the node marked in green as thetrunk point and the node marked in red as one of thecut points. In this example, the purple node will be keptas it has a clear path to the trunk, while the blue nodewill be pruned since its path to the trunk passes throughthe cut point.Our real data for pruning was noisy, with severaloperations taking place at once and no clear definitionas to where cuts were taking place. Instead, we testedour pruning simulator by using the ”SimTreeLS” LiDARsimulator presented by Westling et al. (2020a) to generatepoint clouds for which we can perfectly define removalof matter. Three unique tree models were generated toroughly match the structural characteristics of avocadotrees, and then 4 cut points were specified manually foreach tree, chosen to remove one major limb with eachcut. The matter was removed from the tree in mesh form,where each leaf and limb was visible and no noise waspresent, allowing the matter removal ground truth to begenerated without human error. Different tree candidateswere then generated, whereby each tree was generatedwith between 1 and 4 cuts. Stands of 3 trees each weregenerated by selecting one candidate from each unique tree and placing the trees atop a ground plane in arandomised order at certain inter-tree spacings, from 3mto 8m. 8 replicates were generated at each chosen valuefor spacing, and these 32 stands formed our ground-truthpruned dataset G . For each stand in G , we generatedan unpruned reference stand using the same tree orderand spacing, but a tree model with no cuts, to form ourreference dataset R . We then used SimTreeLS to ”scan”all the data with a virtual LiDAR, including occlusionand sensor noise, to generate the scanned sets L G and L R . For each tree in L R , we computed the pruned matterusing the known cut points and the method describedabove to generate the pruned output set L P . We thencompared the pruned trees from L P to the scannedground truth set L G , and labelled all points accordingto whether they should have been removed or not. Truepositive points were those which were removed in L P and were not present in L G , and so on. We evaluated thesuccess of the pruned matter detection by calculating theF1 score of the points thus labelled. Using our scoring function and pruning effect simulator,we developed a procedure for suggesting which limbsof the tree should be pruned to open up the canopyfor more light. We generate a number of candidatecut locations, simulate the pruning effect, and re-scorethe tree in each new structure. Since thre tree score isbased on raytracing, which can be time consuming, weidentify candidate points by estimating parts of the treewhich contribute towards the global score S . To providethe necessary metadata for this, we first calculate thelight distribution score D i for each voxel as describedin Section 2.1. As when simulating pruning effects, wethen generate a graph by connecting voxels to theirneighbours. We add a ”shade score” to each voxel inthe graph, as illustrated in Figure 5 by counting thenumber of voxels directly below it with a lower D i value.This was designed to identify voxels which were shadingothers, with the assumption that removing these voxelswill open up the centre of the tree to more light.The paths from each node to the trunk of the treeare then generated using A*. For each node, we sumthe number of times it is part of the path to a nodewith a high shade score. Therefore, pruning nodes witha high sum would result in the removal of many shadingnodes. However, since this approach overwhelminglyfavours nodes along the major arteries of the tree (sinceall paths go through them), we divide this score by theproportion of path length between the current node andthe endpoint of the path and call the resultant score j i .We then select candidate points as those in the highest5th percentile of this score. Candidate points are thendown-sampled based on distance to avoid cut pointsin the same location, and the k highest scoring pointsare selected as our pruning set N p . An output of this Fig. 5: Illustration of process for identifying areas ofthe tree which should be removed to improve access tolight. Voxels are scored by counting the number of voxelsbelow them with lower D i scores. Viewing in colourrecommended.process, as well as the scored candidate points, is shownin Figure 6.For each point in N p , we simulate the effect of pruningat that point on the tree and then re-score the tree by ray-tracing over the new point cloud, allowing comparisonsbetween the original score and the new scores.To validate this process, we run pruning suggestionon a virtual tree generated using SimTreeLS as wellas a real tree manually pruned by commercial expertson Simpsons Farms. The real avocado tree used wasscanned following conservative application of selectivelimb removal, with the result showed in Figure 7.For both real and virtual data, we suggest a set N p witha small k value (less than 10), and also present resultswhen all points in N p are used rather than any individualone. ESULTS
Here we present comparisons between our tree scoringfunction and yield, to demonstrate a favourable correla-tion so we can take the scoring function as a reasonabletarget for improving the tree shape.As mentioned in Section 2.1, we scored two sets offruit tree point clouds, one of mature avocado trees andone of young mango trees. When evaluating the mangotrees, we averaged trees within the same replicate andgrowing style to account for outliers, although with theavocado trees we did not have sufficient data points todo this.The results of both comparisons, using the coefficientvalues presented earlier, are shown in Figure 8.
Fig. 6: Illustration of pruning point suggestion. Pointscoloured from blue to red (absolute values unimportant)are candidate points for pruning, namely those nodes inthe graph with the highest j i score. The black points arethe final suggested cut points for this tree.Fig. 7: Visualisation of the manual pruning applied toa particular avocado tree (row-55s-tree-15e). Points inred were removed in the selective limb removal stageof pruning. As described in Section 2.2, we calculated the F1 scoreof simulated matter removal for 137 trees, each with Fig. 8: Comparison of scoring function as a correlatorfor fruit tree yield in avocado and mango trees. Here,”yield” is the total fruit count per tree, and ”weight” isthe average fruit weight in grams. In mango trees, weaveraged the results for trees within the same replicateblock, density and growing style to account for outliers.In avocados, we take each of our 18 data trees individ-ually.a randomly chosen number of cut points and usingvarying inter-tree spacing. Figure 9 shows a qualitativeexample of the output from the pruning operation.Fig. 9: Example output of successful pruning matter sim-ulation. Large blue points are the original cut location,while cyan points are parts of the point cloud identifiedas part of the cut branch. Red points are those estimatedto be removed by the pruning operation.The aggregated quantitative results by number of cutsare presented in Figure 10.
Fig. 10: Results of pruning simulated trees at various treespacings, with a randomly selected number of cut pointsper instance. Scores presented are the average F1 scorefor all instances per spacing and number of cuts.
Tables 2 and 3 present the results for suggestions on avirtual and real tree respectively. For each tree, sevenpoints are suggested, and the score after pruning isreported. We present the change in the overall score aswell as the change in the light distribution D score. Forthe real tree we also present the result following manualcommercial limb removal. Cut D ˜ V ˜ L S D change S changeNone 0.269 1.000 1.000 1.530 0.00% 0.00%A 0.277 0.978 0.969 1.516 3.12% -0.91%B 0.282 0.958 0.947 1.502 4.95% -1.84%C 0.297 0.568 0.595 1.108 10.36% -27.62%D 0.301 0.510 0.537 1.052 12.11% -31.27%E 0.282 0.853 0.931 1.414 4.99% -7.62%F 0.280 0.929 0.905 1.462 4.04% -4.44%G 0.277 0.963 0.940 1.495 2.90% -2.29%H 0.278 0.955 0.942 1.492 3.47% -2.52%I 0.279 0.913 0.904 1.447 3.65% -5.42%All 0.315 0.418 0.420 0.963 16.96% -37.05% TABLE 2: Suggestions for virtual tree. The D score for theentire tree, as well as the normalised values for volumeand total light absorbed are presented. The final score S and measured change in D and S are also presented.Figure 11 shows the qualitative effect of pruning onthe real avocado tree, both with the manual commercialdecision and our pruning decision, specifically cut D. ISCUSSION
The presented framework can, given a LiDAR-scannedpoint cloud of a tree, suggest a number of pruning cuts (a) Unpruned(b) Pruned (manual)(c) Pruned (ours, cut D)
Fig. 11: Visualisation of pruning on real tree. Pointsare coloured by the amount of light absorbed during ayear, and the purple circle illustrates where a limb wasremoved. The ”manual” point cloud is a real LiDARcapture after pruning, while our ”cut D” point cloudwas pruned in simulation based on an autogenerated cutpoint.
Cuts D ˜ V ˜ L S D change S changeNone 0.182 1.000 1.000 1.391 0.00% 0.00%A 0.193 0.955 0.967 1.362 6.02% -2.03%B 0.198 0.916 0.898 1.318 8.84% -5.21%C 0.195 0.948 0.975 1.363 7.33% -1.98%D 0.204 0.889 0.896 1.307 12.50% -6.04%E 0.185 0.987 0.963 1.374 1.66% -1.21%F 0.188 0.956 0.931 1.345 3.39% -3.30%G 0.189 0.949 0.996 1.360 3.80% -2.19%H 0.185 0.978 0.982 1.372 1.59% -1.31%I 0.200 0.880 0.895 1.293 10.16% -7.06%All 0.227 0.761 0.736 1.194 25.15% -14.16%Manual 0.197 0.808 0.989 1.259 8.48% -9.49% TABLE 3: Suggestions for real tree. The D score for theentire tree, as well as the normalised values for volumeand total light absorbed are presented. The final score S and measured change in D and S are also presented.for selective limb removal and then report the resultingimprovement according to a defined scoring function.Figure 8 presents a comparison between S scores andyield characteristics. While the correlations are far fromconclusive, and fruit weight for mangoes in particularhas no discernible correlation, there is evidence of a linkbetween the score and the yield. ¨Orn (2016) suggestedthat the volume of the tree is the most important prop-erty for predicting fruit count, while the total light ab-sorbed is better for the average fruit weight. Consideringthe observations of Stassen et al. (1999) and Marini et al.(2020), the D score is likely to be related to the sugarand oil content of the fruit produced in different parts ofthe canopy, with higher D scores correlating with betterquality fruit in more parts of the canopy. However, thatis only speculation until it can be experimentally veri-fied. Due to the complexity of judging tree quality andgrowth, our scoring function and coefficients, reflected inthe yield graphs presented, are merely a starting pointfor the evolution of our pruning suggestion framework,to be customised depending on application.The results presented in Figure 10 shows that thepruning simulator generally works well, with an averageF1 score of 0.78. It can be observed that the results weregenerally better when the spacing is larger, as couldbe expected. When the tree spacing is only 3m, thereis significant overlap between adjacent trees, thoughthe graph-based operation used is designed to performreasonably well in that context. The results also suggestthe method performs slightly better with more cutsrather than fewer. This could be caused by overlappingpredictions on branches which are close together.Results shown in Table 3 demonstrate how our totalsuggestion procedure performs on a real avocado treefor which we also have the actual commercial pruningdecision. One thing to note is that the change in total S score is always negative to varying degrees. This islikely due to the importance of the tree volume andtotal light absorbed, which are always reduced whenmatter is removed. The premise of pruning is that somesacrifice in volume is required to promote regrowth and improve light distribution, though this may be adeficit in our selection of coefficients. Since we selectedthese by comparing to yield, the importance of lightdistribution (which has a smaller impact on yield thanvolume, as noted by ¨Orn (2016)) is reduced. Under theassumption that improved light distribution has otherbenefits, we also presented the change in D score asindependent from the volume V and total light L , andhere we see the inverse effect - for all of our suggestedcut points, as well as the manual limb removal, the Dscore improves after pruning. In the case of three of oursuggestions (B, D and I), the tree ends up with a higher D score after pruning with a higher total score(due toless matter removed) when compared to the manuallyselected limb. Cut D in particular increases D by 4% and S by 3.45% over the manual cut, and the result of thiscut is shown in Figure 11. While the figure is difficultto grasp in two dimensions, our method qualitativelyopens up the canopy more than the manual cut, as canbe seen by the reduction in dark blue points (whichtend to be shaded during the year) and the scatteringof more red and orange points throughout the lowercanopy. This improved distribution comes without muchphysical volume being removed.Table 2 presents similar results for our suggestionsystem on a virtual, computer-generated tree. The samepatterns can be seen here, with volume leading to aconsistently negative change in S , but with the cleanerscan generated by simulation, more significant cuts wereimplemented with volumes up to 50% of the tree.These results, in particular the ones for the real tree,suggest the potential of our procedure to suggest prun-ing locations which can open up the canopy more whileremoving less matter than current commercial experts,though of course much more experimental validationwould be required to confirm this.While existing simulation-based approaches requirecomputer-generated trees (like those generated by Whiteand Hanan (2016)) or high-quality captures (like Arika-pudi et al. (2015)), our method was developed to workwith low-quality, high-noise LiDAR captures like thoseproduced by handheld (Westling et al. (2018); Bosse et al.(2012)), mobile (Underwood et al. (2016)), or even aerial(Wu et al. (2020)) equipment.This framework overall provides a potential startingpoint for developing tools for commercial yield improve-ments, or even automated pruning. Naturally, the param-eters used can be tweaked to fit various contexts, withthe basic system being appropriate for any tree struc-ture with a desire for uniform light distribution whilebalancing tree volume. Other pruning objectives canalso be included without much adjustment, for instancethe height-limiting approach investigated by Thorp andStowell (2001) to move the fruiting area closer to theground for operational reasons. The individual parts ofthe framework can also be extracted or shuffled for otherpurposes. One example of this could be in developingtraining systems like that presented by Kolmaniˇc et al. (2017), wherein a human operator suggests a cut location(taking the place of our pruning suggestion system) andis provided feedback on how well that decision wouldopen the canopy. Also, the pruning effect simulationcould easily be adapted to perform hedging operationsrather than selective limb removal, allowing the frame-work to perform that task as well. As mentioned earlier, the scoring function we developedis fairly simple and does not take into considerationmany parameters of fruit tree quality or attractiveness. Tothis end, further experimentation should be undertakento validate in particular that the D score component doesserve as a predictor for higher quality fruits. This wouldrequire point cloud scanning of fruit trees, paired with adetailed survey of the yield to capture fruit quality, sizeand sugar/oil content in different parts of the canopy.Other parameters can also be derived from the pointcloud to add to the S score. One example could be usingmethods like those presented by Vicari et al. (2019) andWestling et al. (2020c) to segment the leaf and trunkmatter, to include properties like the woody-to-total arearatio described by Ma et al. (2016).The suggestion method itself could also be furtherdeveloped using more sophisticated gradient descentapproaches, or a more expansive optimisation technique.The approach suggested here is also simplistic in itscomputation of shade score, assuming that each voxelonly shades those directly below it, while the truth couldbe more accurately modelled as a cone of influence, withparameters inferred by the integrated sky.With further development, this framework could alsoinclude regrowth simulation to more accurately modelthe long-term effects of pruning, as presented by Strnadet al. (2020). This would require some careful interpo-lation, given that methods like these typically require ahigher fidelity tree model than can be captured in reason-able time using currently available LiDAR technology. ONCLUSION
We have presented a framework for suggesting strategiesfor selective limb removal pruning on fruit trees, witha specific focus on avocados and mangoes. Our resultsshow correlation between the scoring function we devel-oped and yield, and we are able to successfully simulatethe removal of plant matter given specific cut points. Bydynamically suggesting cut points in order to removetree matter which is casting shade on parts of the tree,we also presented a method for suggesting cuts whichimproved the light distribution through the canopy byup to 25.15%. The steps in this framework can be furtherdeveloped to provide a pruning suggestion system withcommercial applicability.
Acknowledgements
This work is supported by the Australian Cen-tre for Field Robotics (ACFR) at The Universityof Sydney. For more information about robots andsystems for agriculture at the ACFR, please visithttp://sydney.edu.au/acfr/agriculture. R EFERENCES
R. Arikapudi, S. Vougioukas, and T. Saracoglu. Orchardtree digitization for structural-geometrical modeling.In
Precision agriculture’15 , pages 161–168. WageningenAcademic Publishers, 2015.Australian Centre for Field Robotics (ACFR). Commaand snark: generic c++ libraries and utilities forrobotics. https://github.com/acfr/, 2012. Accessed:2017-02-18.M. Bosse, R. Zlot, and P. Flick. Zebedee: Design of aspring-mounted 3-d range sensor with application tomobile mapping.
IEEE Transactions on Robotics , 28(5):1104–1119, 2012.D. Da Silva, L. Han, and E. Costes. Light interceptionefficiency of apple trees: a multiscale computationalstudy based on mapplet.
Ecological Modelling , 290:45–53, 2014.J. Estornell, L. A. Ruiz, B. Vel´azquez-Mart´ı, I. L ´opez-Cort´es, D. Salazar, and A. Fern´andez-Sarr´ıa. Estima-tion of pruning biomass of olive trees using airbornediscrete-return LiDAR data.
Biomass and Bioenergy , 81:1–7, 2015. ISSN 09619534. doi: 10.1016/j.biombioe.2015.07.015.S. Kolmaniˇc, S. Tojnko, T. Unuk, and ˇS. Kohek. Thecomputer-aided teaching of apple tree pruning andtraining.
Computer Applications in Engineering Educa-tion , 25(4):568–577, 2017.L. Ma, G. Zheng, J. U. Eitel, T. S. Magney, and L. M.Moskal. Determining woody-to-total area ratio usingterrestrial laser scanning (TLS).
Agricultural and ForestMeteorology , 228-229:217–228, 2016. ISSN 01681923. doi:10.1016/j.agrformet.2016.06.021. URL http://dx.doi.org/10.1016/j.agrformet.2016.06.021.R. P. Marini, S. Sherif, A. H. Smith, et al. Physiology ofpruning fruit trees. 2020.C. M. Menzel and M. Le Lagadec. Increasing theproductivity of avocado orchards using high-densityplantings: a review.
Scientia Horticulturae
Yearbook of the California Avocado Society ,44:42–4, 1960.S. ¨Orn.
Estimating Light Interception of Orchard Trees UsingLiDAR and Solar Models jsf?dswid=8933 { & } pid=diva2 { % } { & } c=1 { & } searchType=SIMPLE { & } language=en { & } query=samuel+orn { & } af= { % } { % } { & } aq= { % } { % } { % } { % } { & } aq2= { % } { % } { % } { % } { & } aqe= { % } { % } { & } noOfRows=50 { & } sortOrder=author { } sort { } asc { & } onlyFullText=false { & } sf=all.G. J. Partida Jr. Avocado canopy management for greateryields and orchard efficiency. California Avocado SocietyYearbook , 80:117–131, 1996.L. Purcell. Tree pruning essentials.
Forestry and NaturalResources. Purdue Extension , 2015.H. Sinoquet, G. Sonohat, A.-M. Potel, P. Monney, andP.-E. Lauri. Using virtual plants to estimate light dis-tribution at intra-canopy scale in apple trees: Methodpresentation and assessment. In
VIII InternationalSymposium on Modelling in Fruit Research and OrchardManagement 803 , pages 225–234, 2007.B. Snijder and P. Stassen. Strategies for renewal ofunproductive older avocado orchards with severe en-croachment problems.
South African Avocado Growers’Association Yearbook , 18:56–58, 1995.P. Stassen, B. Snijder, and Z. Bard. Results obtainedby pruning overcrowded avocado orchards.
RevistaChapingo Serie Horticultura , 5:165–171, 1999.D. Strnad, ˇS. Kohek, B. Benes, S. Kolmaniˇc, and B. ˇZalik.A framework for multi-objective optimization of vir-tual tree pruning based on growth simulation.
ExpertSystems with Applications , 162:113792, 2020.A. Tagarakis, S. Koundouras, S. Fountas, and T. Gemtos.Evaluation of the use of lidar laser scanner to mappruning wood in vineyards and its potential for man-agement zones delineation.
Precision Agriculture , 19(2):334–347, 2018.L. Tang, C. Hou, H. Huang, C. Chen, J. Zou, and D. Lin.Light interception efficiency analysis based on three-dimensional peach canopy models.
Ecological Informat-ics , 30:60–67, 2015.O. Tange. Gnu parallel - the command-line power tool. ;login: The USENIX Magazine
HortScience , 36(4):699–702, 2001.J. P. Underwood, C. Hung, B. Whelan, and S. Sukkarieh.Mapping almond orchard canopy volume, flowers,fruit and yield using lidar and vision sensors.
Com-puters and Electronics in Agriculture , 130:83–96, 2016.M. B. Vicari, M. Disney, P. Wilkes, A. Burt, K. Calders,and W. Woodgate. Leaf and wood classification frame-work for terrestrial lidar point clouds.
Methods inEcology and Evolution , 10(5):680–694, 2019.F. Westling. Avocado tree point clouds before and afterpruning. Mendeley Data.F. Westling, J. Underwood, and S. ¨Orn. Light interceptionmodelling using unstructured lidar data in avocado orchards.
Computers and electronics in agriculture , 153:177–187, 2018.F. Westling, M. Bryson, and J. Underwood. Simtreels:Simulating aerial and terrestrial laser scans of trees. arXiv preprint arXiv:2011.11954 , 2020a.F. Westling, K. Mahmud, J. Underwood, and I. Bally.Replacing traditional light measurement with lidarbased methods in orchards.
Computers and Electronicsin Agriculture , 179:105798, 2020b.F. Westling, D. J. Underwood, and D. M. Bryson.Graph-based methods for analyzing orchard tree struc-ture using noisy point cloud data. arXiv preprintarXiv:2009.13727 , 2020c.N. White and J. Hanan. Use of Functional-StructuralPlant Modelling in Horticulture. 2012.N. White and J. Hanan. A model of macadamia withapplication to pruning in orchards.
Acta Horticulturae ,1109(June):75–81, 2016. ISSN 05677572. doi: 10.17660/ActaHortic.2016.1109.12.J. D. Wilkie, J. Conway, J. Griffin, and H. Toegel. Relation-ships between canopy size, light interception and pro-ductivity in conventional avocado planting systems.
The Journal of Horticultural Science and Biotechnology , 94(4):481–487, 2019.D. Wu, K. Johansen, S. Phinn, and A. Robson. Suitabilityof airborne and terrestrial laser scanning for mappingtree crop structural metrics for improved orchard man-agement.
Remote Sensing , 12(10):1647, 2020.W. Xiao, S. Xu, S. O. Elberink, and G. Vosselman. Changedetection of trees in urban areas using multi-temporalairborne lidar point clouds. In
Remote Sensing of theOcean, Sea Ice, Coastal Waters, and Large Water Regions2012 , volume 8532, page 853207. International Societyfor Optics and Photonics, 2012.W.-W. Yang, X.-L. Chen, M. Saudreau, X.-Y. Zhang, M.-R.Zhang, H.-K. Liu, E. Costes, and M.-Y. Han. Canopystructure and light interception partitioning amongshoots estimated from virtual trees: comparison be-tween apple cultivars grown on different interstockson the chinese loess plateau.