SwarmFarmRobotics：2022年為澳大利亞蘋果行業開發農業技術解決方案最終報告上（英文版）（68頁）.pdf

《SwarmFarmRobotics：2022年為澳大利亞蘋果行業開發農業技術解決方案最終報告上（英文版）（68頁）.pdf》由會員分享，可在線閱讀，更多相關《SwarmFarmRobotics：2022年為澳大利亞蘋果行業開發農業技術解決方案最終報告上（英文版）（68頁）.pdf（68頁珍藏版）》請在三個皮匠報告上搜索。

1、 Final Report Developing Agri-Tech Solutions for the Australian Apple Industry Project leader:Andrew Bate Delivery partner:SwarmFarm Robotics Project code:AP16005 Hort Innovation Final Report:Developing Agri-Tech Solutions for the Australian Apple Industry(AP16005)Project:Developing Agri-Tech Soluti

2、ons for the Australian Apple Industry(AP16005)Disclaimer:Horticulture Innovation Australia Limited(Hort Innovation)makes no representations and expressly disclaims all warranties(to the extent permitted by law)about the accuracy,completeness,or currency of information in this Final Report.Users of t

3、his Final Report should take independent action to confirm any information in this Final Report before relying on that information in any way.Reliance on any information provided by Hort Innovation is entirely at your own risk.Hort Innovation is not responsible for,and will not be liable for,any los

4、s,damage,claim,expense,cost(including legal costs)or other liability arising in any way(including from Hort Innovation or any other persons negligence or otherwise)from your use or non-use of the Final Report or from reliance on information contained in the Final Report or that Hort Innovation provi

5、des to you by any other means.Funding statement:This project has been funded by Hort Innovation,using the apple and pear research and development levy and contributions from the Australian Government.Hort Innovation is the grower-owned,not-for-profit research and development corporation for Australi

6、an horticulture.Publishing details:ISBN 978-0-7341-4799-8 Published and distributed by:Hort Innovation Level 7 141 Walker Street North Sydney NSW 2060 Telephone:(02)8295 2300 .au Copyright 2022 Horticulture Innovation Australia Final report-AP16005 3 Hort Innovation Contents Public summary.4 Keyword

7、s.5 Introduction.6 Methodology.7 Variable Timing of Application Sprayer.7 Phase 1-Hardware Design.7 Phase 2-Individual tree testing.7 Phase 3-Flow rate validation.8 Phase 4-Orchard testing.9 Apple Snapper System.10 Hardware Selection.10 Software Development.11 Training Pipeline.11 Results.12 Variabl

8、e Rate Sprayer.12 Apple Snapper System.12 Image Detection Models.12 Outputs.13 Outcomes.19 Recommendations.23 Intellectual property.24 Appendices.25 Final report-AP16005 4 Hort Innovation Public summary To develop Agri-Tech solutions for Horticulture,and specifically to the Apple Industry,by providi

9、ng digital“mapping”of both flower stage and canopy development.The objective was to bring a product to market that will perform autonomous flower stage mapping and tree canopy measurement and to provide decision support for variable timing spraying specifically for primary chemical thinning.This can

10、 then be utilised by growers to make better decisions around orchard management including quality and yield improvements,reduce production costs including labour and chemical use,and increase consistency across all trees.This 3-year project had as its objective to develop a product to usher in the n

11、ew digital age of digital agriculture in the Apple Industry.This report has been provided for Hort Innovation Australia under the terms of the research agreement.Final report-AP16005 5 Hort Innovation Keywords Variable rate spraying;variable timing of application;VTA;decision support tool;apple snap

12、per;machine learning;live processing;flower stage;flowering stage detection;canopy density.Final report-AP16005 6 Hort Innovation Introduction The objective of this project was to develop a product that will perform flower stage mapping and tree canopy measurement and to provide decision support for

13、 variable timing of application spraying.This project helped to usher a new digital age of digital agriculture in the apple industry.This project was led by SwarmFarm Robotics with both University of NSW and ADAMA being partners in research and development of the technology.A key output of this proj

14、ect is a system complete and presented ready for commercialisation.The system has been validated in-orchard and ready for industry adoption though commercial partnerships.This project has successfully proven:1.Live processing of flower stage/distribution images at 10km/hr 2.Instantaneous upload of d

15、ata to the decision support tool website upon block scan completion 3.Growers are able to manipulate the latest data on decision support tool website to obtain an understanding of the orchards variability and growth stage 4.Growers are able to create prescription spraymaps via the web interface as t

16、he scan is finished.5.Growers are able to load a spray map onto the Raven controller and spray a block minutes after finishing a scan 6.Accurate execution of the spraymap on a tree by tree basis 7.The tops and bottoms of trees can be sprayed at different rate 8.The final product can be retro fitted

17、onto any existing machinery,such as tractors,side by sides or quad bikes.Final report-AP16005 7 Hort Innovation Methodology Variable Timing of Application Sprayer SwarmFarm worked with industry to develop an orchard sprayer that utilises precision agriculture technology to apply primary thinners,spe

18、cifically ATS,on a tree by tree basis.Feedback from the precision ag industry after the initial design(focusing on variable rate spraying on a tree by tree basis)concluded vastly changing the application rate was not achievable at the accuracy level desired and also resulted in off-label application

19、.After discussion with industry,the Variable Timing of Application concept was created.This focused on targeting individual trees at the correct flowering stage with either the full rate,or nothing.SwarmFarm worked with the industry to develop six objectives to determine the feasibility of tree by t

20、ree application:1.The ability to switch nozzles on/off over 1 metre tree spacings 2.Accurate targeting of specific trees 3.Maintain pressure/rate while targeting each tree 4.Even coverage over the tree(spray pattern)5.Correct droplet size 6.The ability to spray up to 8km/h To test these objectives f

21、our phases were established and field trials were carried out.Phase 1-Hardware Design When the project kicked off,the first step was to design the sprayer hardware and determine if it would be possible to actually target individual trees with off the shelf hardware.Silvan Australia were contacted an

22、d a collaborative partnership was formed with the aim to deliver a sprayer to industry that could accurately target individual trees.There were two sprayer options;a tower sprayer or an airblast sprayer.After discussions with Silvan and growers in the industry,the decision to focus on tower sprayers

23、 was made,with the belief that it would be best for spray delivery to target.Ravens Rate Control Module and Sidekick pro pumps were used as the control system on the tower sprayer.This was the best technology available that could change the application rate quickly,a requirement for tree by tree spr

24、aying.However,after testing in the orchard,it was found to take around 12 m long sections(distance up the orchard row,roughly 8 trees)to hit the target rate.Meaning tree by tree variable rate was not possible with this design.Feedback from industry concluded that spraying an individual tree based on

25、 its flowering stage was the most effective primary thinning method.So,throughout 2020 the tower sprayer was fitted out with Raven Hawkeye control equipment under the recommendation and guidance of Matt Daniels from Raven Industries.The tower sprayer performed well in all our testing however due to

26、the Covid 19 situation and the low hanging top wires in Batlow we were unable to test the tower sprayer in an orchard.The feedback from growers in the Batlow and Shepparton areas concluded that uptake would be low with the expensive new tower sprayer design,and that a standard airblast sprayer that

27、could do the same job would have the best chance at being adopted by growers.So,for the 2021 flowering season a new airblast sprayer was purchased from Silvan and kitted out with the Raven Hawkeye control equipment.The following tests were conducted using this sprayer.Phase 2-Individual tree testing

28、 To ensure the first 2 goals around accuracy and timing of the chemical application were achievable in an orchard,significant amounts of testing was carried out with both the tower sprayer and airblast sprayer equipped with the Raven Hawkeye control system.The main test involved mapping a series of

29、posts location(spaced 1 metre apart)and creating spray maps based on these locations.These posts mimic trees planted 1 metre apart in an orchard.The aim was to prove the sprayer could switch on/off to target every second tree.To visualise this,butcher paper was rolled out over the posts and dye was

30、added to the spray vat.This allowed us to see exactly where the sprayer was delivering spray,compared to where it was targeting.Final report-AP16005 8 Hort Innovation Figure 1.Initial testing setup at Harden NSW The tractor was driven 2 metres from the posts with the prescription map loaded to targe

31、t every second spacing.Multiple passes were conducted on fresh paper with different configurations in the Raven Controller until the spray application was hitting the desired target positions.Figure 2 shows the final results from this test.You can see that every second spacing in the image is colour

32、ed pink from the die in the tank.This test was conducted at 8km/hr,giving confidence the sprayer system could operate in an orchard,targeting every second tree at the desired speed.It also demonstrated that we could fully turn on and off every nozzle every metre spacing which we determined would be

33、as narrow as possible while still giving adequate coverage to the on sections without over-spraying the off sections.From the prescription maps generated during this season,it was noted that very rarely does the sprayer encounter a row where every second tree is much more or less advanced.Usually,th

34、e trees are more consistent with a few outliers.This means the sprayer will not have to perform the amount of switching we demonstrated during testing.Figure 2.Results from fence line testing in Harden NSW Phase 3-Flow rate validation The next tests carried out ensured that the flow rate from each n

35、ozzle was correct.This was completed in November 2020.A hose was placed over a nozzle at random which then directed the flow out of the nozzle to a bottle.The volume in the bottle was then compared to the applied volume that should have been sprayed according to the prescription map.The flow rate te

36、sts were successful,proving application rate was correct even when switching nozzles on/off at close intervals.Final report-AP16005 9 Hort Innovation Figure 3.An Example of nozzle flow rate testing.source:https:/ 4-Orchard testing Phase 2 and phase 3 testing gave confidence that variable timing of a

37、pplication spraying was achievable within an orchard.The configuration of the sprayer was tried and tested,the next stage was to determine the accuracy and coverage when operating within the orchard row.Several trials were completed in consultation with Sally Bound,Senior Research Fellow at Tasmania

38、n Institute of Agriculture,agronomists,and Growers in the Batlow area.These trials were focused on droplet size,coverage and penetration into the foliage at different speeds.Dye was used to test the overall coverage as well as droplet size and overspray along an orchard row.After promising results w

39、ith the dye,water sensitive paper was then used to conduct more accurate and measurable tests.Figure 4.The sprayer spraying the top left and bottom right sections of trees Much like the fence line tests,each individual trees GPS location was mapped.Shapefiles were output of these locations and a pre

40、scription map was generated to spray each individual tree and not spray the rest of the block.Water sensitive paper was then placed at different distances from the mapped location to determine the accuracy of the Raven system,when targeting an individual tree.It was then only a matter of trialling d

41、ifferent speeds to ensure consistent results were found in the orchard that matched the accuracy to the paper fence line test.Once the accuracy of the system was validated,focus turned to the coverage of the sprayer.Water sensitive paper was placed in the rows behind the tree being targeted.These te

42、sts are outlined in detail in Appendix 1.Final report-AP16005 10 Hort Innovation Apple Snapper System The apple snapper system describes the complete data collection pipeline from capturing images,processing with flower stage/intensity and canopy density algorithms,automatic upload,to viewing on the

43、 decision support tool webpage and outputting an application spray map.The steps of the process as follows.Data collection/scanning of trees.Live image processing.Automatic upload and storage.Decision Support Tool Website to help the Grower visualise the orchard.Generating the prescription spray map

44、 from the data.Figure 5:Full system pipeline developed from data collection to spraying with application maps The final data collection system based on feedback from growers and the latest research in agricultural machine vision and automation was developed and fine-tuned over 3 seasons in orchards.

45、The final system consists of multiple hardware and software modules working together to make a seamless process from data collection to spray map output.Outlined below is the methodology and justification of the research and verification done to produce the final product.Hardware Selection Hardware

46、selection was considered first,with different cameras,LiDARs(light and detection ranging),GPS,IMU(inertial measurement unit),computers tested and evaluated to get the best results possible.Cameras were chosen on the basis of how much detail was needed to correctly identify all stages of the flowerin

47、g life-cycle,from pink tip through the petal fall.It was important for early stage estimation to have a high enough resolution to be able to see the tip/bud development.Different wide angle lenses were sourced and tested to ensure the whole tree,in different common row widths,could be captured with

48、minimal distortion.Final costs were reduced as testing showed that only one camera per side was needed to capture the full tree,thus eliminating the need for multiple cameras and complex software for syncing the images together.The collection rigs location in the orchard was a critical component so

49、significant time was taken testing multiple systems to achieve the best accuracy.These tests involved driving through the orchard and plotting the positions of what from different GPS receiver systems.The final GPS has an integrated IMU with its own filters so there are no drop-outs or jumps in the

50、data when driving under trees.During the early development of the image detection models,the images were captured and then sent to UNSW for processing.Once the Image detection models were built and validated,focus moved to enabling local processing of images to eliminate the need to upload and send

51、images from the field to UNSW for the processed output.Edge AI computers have become very powerful and energy efficient due to the research for autonomous cars and robotics,which need highly efficient powerful computers to process the many connected safety devices.This technology enabled us to shift

52、 to the ability to test live processing of images,without the need for large power sources.The final platform used was the Jetson Xavier SOM with a Neousys Carrier board that expanded the number of ports for the cameras,GPS,LiDAR and 4G/Wifi Modem.Initially a large lattice structure was used to test

53、 numerous cameras and lighting arrays.However,once the camera and lens selection were finalised the structural design was focused around their ideal position.Next,many flood lights were used to capture consistent images which had to be adjustable for different settings,however in the final season st

54、robe lights allowed the structure to be much more compact.Final report-AP16005 11 Hort Innovation Software Development The software underwent many different iterations throughout the project.Early on it enabled quick turnaround testing of new components and collection methods but was not optimised f

55、or a commercial product.Once the final hardware was selected,work was done to test live processing and capture.The specific AI Computer we chose allowed us to optimise the image detection models to run much faster on the selected hardware enabling live processing for each tree at 8km/h.The system ca

56、n also be configured to process every second tree in the orchard,to give a lower resolution map of the orchard.Speeds of up to 30km/hr can be achieved when only processing every second tree.The final software stack underwent a large refactor in the final year of the project to increase the robustnes

57、s and error handling ability required to allow for operation by growers.The code was also stripped back and modularized to make maintenance straightforward.The user interface was not developed until the final season due to regular changes in internal software communication formats during development

58、.Once the hardware,inputs,and outputs were finalised,a user-friendly user interface with the ability to control and debug all aspects of the collection rig was developed.Connected system To make the process from capture to producing a spray map as seamless as possible,a full pipeline was implemented

59、 to push data from the Apple Snapper to a cloud service.This means growers do not need to manually handle or sync data.To get the user started as quickly as possible,boundaries for blocks can be created on a map that is loaded automatically to the collection rig.The user then inputs the required set

60、tings such as row and image spacing and it is then a one button press to start recording.The user can continue to drive as many blocks they want until finished and then can confirm the data collected and upload straight to the cloud.Once uploaded the maps are available immediately in the decision su

61、pport tool website.The Decision support tool development was not completed until the final hardware,software,and file formats were decided on.This allowed us the freedom of not being locked into a single format early on and the ability to select the best format based on a complete set of information

62、.The key part to validating the usability and value in the decision support tool was being able to represent and manipulate the data in an intuitive way to the growers,allowing them to make the most informed spray decisions.Work went into informing and showing growers the data collected in many diff

63、erent formats.Feedback from these discussions contained important insights into the information the grower wants to see to make the best informed decisions for their orchard.An overview of the features available on the Website are expanded upon in Appendix 5.Training Pipeline A quick turnaround trai

64、ning pipeline was developed to quickly re-train the algorithm models as new data sets were captured during the season.40 quads(0.5m X 0.5m)were placed in trees and were counted/classified three times a week throughout the season.The location of each quad was known to the Collection Rig so images wit

65、h quads were able to be captured separately and sent through to UNSW when scanning was complete.All flowering buds were then counted and classified by an agronomist and were passed onto UNSW who were then able to quickly process the new images and counts to improve the imaging algorithm models.This

66、streamlined process allowed us to rapidly get more accurate maps as the season progressed.Some of the quads were also used as validation and accuracy indicators to see how much improvement we were getting with more trained images.As the quads were consistently at the same height across the block,the

67、y did not represent the full variability of an apple tree.Therefore,larger full sized quads were imaged and counted to be used as additional validation when analysing the trained models.Throughout the flowering period we walked through orchards in the Batlow area with growers and agronomists.During

68、these walks we discussed thinning decisions on a tree-by-tree basis.The location of these trees was then brought up on the decision support tool,with the image of the location being matched to the tree we were looking at.This was used to improve grower trust in the system as they could use their exp

69、erience and judgement to guess what stage the tree is at themselves,and whether they would want to spray it.We could compare the growers summary of the tree to the data we had recorded,and after going through this process for a few trees their confidence in the system improved drastically.Final repo

70、rt-AP16005 12 Hort Innovation Results Variable Rate Sprayer Variable timing of application and variable rate spraying were achieved in this project.Growers can utilise the spray technology developed in this project to spray specific trees within an orchard at a particular growth stage(i.e.70%king bl

71、oom and above)with a particular rate(1000L/ha).This has been proven to be accurate and effective.An industry know-how guide has been developed to give growers and spray manufacturers the technical details on how to configure any Airblast or Tower sprayer to have variable timing and variable rate cap

72、abilities.The sprayer was able to:Action the user-defined prescription map to deliver variable rate and variable timing of spray application Accurately target spray applications independently between top and bottom sections of the canopy Track coverage area and record as-applied maps for regulatory

73、documentation Current product features:Variable Rate Application-spray the top and bottom half of the tree with different water rates,as defined by the DST user.Section Control actioning pre-defined prescription map to only spray those trees that are at the correct flowering stage(VTA Variable Timin

74、g of Application)Output at the completion of a spray application,an as-applied spray application map Both the airblast and tower sprayer performed extremely well throughout the years of testing.Feedback from industry concluded that updating current airblast sprayers is the most likely avenue for bro

75、ad adoption.The tower sprayer may be more accurate,however it is not suitable in orchards with low wires and it has a far greater cost.It is recommended that farmers maintain their current airblast sprayer,and attach the kit to it,thus reducing capital costs of the technology adoption.This would be

76、a huge improvement to production and productivity within the industry.Apple Snapper System For the final product of the project,focus was directed on producing a commercially rounded product that demonstrated the full features developed in this project.Some of the key areas that were prioritised inc

77、luded:Optimisation and streamlining the software stack Creating a more user-friendly interface for controlling and monitoring the Collection Rig Automatic upload and storage of images and produced maps Redesign the Collection rig to a more compact unit Integrate strobe lights to enable faster collec

78、tion All these solutions were developed in time for the 2021 flowering season and only minor adjustments were needed during the season.Data was collected in the Batlow region over 6 weeks across 10 blocks.The two collection rigs recorded over 150 hours each with over 1.5 million images captured and

79、processed in real-time.All the processed maps and a subset of images were uploaded to the cloud and viewable immediately on the Decision Support Tool Website.The details of the improvements and high-level details of the final product is located in Appendix 4 Image Detection Models The resulting data

80、 collection system centres around the image detection models developed by UNSW which output the intensity and stage data extracted from the images.UNSWs report will go into more detail into the methodology and results acquired in the development process.The final models allow for the detection of th

81、e 8 stages of an apple flower from green tip all the way through to petal fall.Detecting each stage gives valuable information on each individual tree and its stage in the flowering process to the grower,allowing for more targeted decisions around the timing of application.Final report-AP16005 13 Ho

82、rt Innovation Outputs Table 1.Output summary Output Description Detail Magazine articles and written communication There were six articles published throughout the project,providing updates to industry.The articles were published in industry magazines,including APAL magazines,Fruit Grower Victoria p

83、ublications.The URLs are provided below https:/apal.org.au/agritech-to-help-with-thinning/https:/ Orchard walks The future orchard walk program was used to promote the work done in the project and update growers in the industry.The first two years project staff attended and presented at all future o

84、rchard walks.However,with the COVID restrictions of the last two years this has not been possible.Annual conference(Hort Innovation)Attend the Hort Innovation conference and provide an update to the industry.All Hort Innovation conferences held throughout the project were attended.We were unable to

85、obtain a speaking position at the conferences.SwarmFarm were able to utilise the ADAMA conference stand to show the robot and generate discussion and excitement amongst growers.Many growers took the opportunity to talk to project staff individually.Annual video summary Create a video to summarise th

86、e yearly development and promote the project to industry.Throughout this project 8 video updates have been recorded.3 have been yearly updates to show progress of the project.The last 4 videos are explainers to assist in adoption and update of the final product.A playlist of videos has been created,

87、these can be found at the following link.https:/ Industry guide for sprayer The industry guide outlines in detail how to build a variable rate sprayer.SwarmFarm has utilised off the shelf components from the broadacre industry to develop a variable timing of application spray system/solution that ca

88、n be fitted to any tower or airblast sprayer.The commercialisation path for the sprayer is via an industry know how guide(appendix 6).The intention of this guide is to explain in details:Final report-AP16005 14 Hort Innovation Hardware component list Hardware installation guide Sprayer configuration

89、 Importing and utilising prescription maps This guide will be distributed via industry channels to growers and spray manufacturers for widespread adoption.The hardware used in this project has been sourced from Raven Industries and is readily available and supported.Similarly,there are competitors t

90、o the Raven control system that are able to modify the setup of their existing spray hardware and computers,to deliver the same functionality.This is ideal for the industry and should lead to multiple manufacturers offering competitive systems to Australian growers.GIS protocol for orchard mapping f

91、or autonomous operation A GIS convention was developed throughout this project for georeferencing an orchard.This is used by both the flower stage estimation platform as well as the SwarmBot system End row posts were surveyed with RTK GPS as well as an overall boundary polygon for each block.The row

92、 end locations were used to generate obstacle zones around every row of trees in the orchard.The SwarmBot used these obstacle boundaries to make sure it never generated a path that hit or damaged a tree.Localisation solution for orchard operation Ability for farm vehicles to accurately localise(unde

93、rstand its position in the global environment)within an Apple Orchard so precision agriculture practises can be utilised Precision agriculture utilises highly precise GPS positioning for its field practises.In broadacre this includes,variable rate maps,auto steer and many more field applications.GPS

94、 and precision agriculture is currently not commonly used in the Apple Industry.GPS is an absolute global reference to a position on the ground.SwarmFarm recognised the benefits of using GPS technology for both the data collection rig as well as the sprayer.The benefit of this is all data is referen

95、ced off the same positions.There is no need to estimate/localise off tree trunks etc.The reason GPS is not commonly used in orchard environments is due to the canopy shadowing the receiver,causing intermittent dropouts.SwarmFarm worked with a third party company to test and improve firmware for IMU+

96、GNSS fused GPS hardware,with the final solution being a robust GPS system for orchard applications.Communication solution for orchard operation The method of communication within an orchard is dependent on many factors.Including orchard location,proximity to 4G towers,orchard density/spacing,tree he

97、ight and antenna configuration on-board the SwarmBot.SwarmFarm has partnered with various communication providers to trial different solutions for machine communication in challenging environments.Trials that have been carried out include:1.4G link to service provider i.e.,Telstra/Optus 2.Local LTE

98、tower on farm-essentially a local 4G tower with a backhaul link to the internet 3.Multi-carrier systems-Multiple models that switch Final report-AP16005 15 Hort Innovation between the various carriers in real time to maintain optimum bandwidth 4.Satellite link via Iridium network 5.Local Wi-Fi meshi

99、ng network consisting of various local nodes,connected via a backhaul link to the internet Due to the close proximity of a Telstra tower in the Batlow region,SwarmBot Papa was able to continuously communicate over 4G while operating autonomously,allowing for consistent communication to connected dev

100、ices.The Wi-Fi was also demonstrated to be an effective communication mode when operating nearby.As both of these options were sufficient in providing maximum safety and operation control,no other communication solution was required.However,for orchards that do not have a stable 4G link,there are mu

101、ltiple communication solutions that can be tailored to fit.SwarmFarm recommends a private LTE networks as the best solution for farms outside of good cellular coverage areas.Safe obstacle detection system for autonomous operation in an orchard SwarmFarm has developed early stage algorithms for safe

102、obstacle detection in orchard environments.The development and work undertaken in this area was completed outside of this project,however we were able to utilise the technology to demonstrate safe obstacle detection within an orchard.There are two standalone obstacle detection systems that SwarmFarm

103、 utilised for obstacle detection within an orchard.1.LiDAR/Time-of-flight sensors are used to detect large obstacles as well as determining the traversability of the upcoming terrain.This system was originally developed for broadacre applications,but was able to be modified to operate in orchards by

104、 filtering tree limbs that overhang the row.This method is still under development and currently has a high count of false positive detections.However,we are certain this can be improved with further development.2.Human Detection system utilising machine learning is another obstacle detection system

105、 used in orchard environments.By using machine learning methodologies,we were able to accurately detect humans and if a person stands near a SwarmBot it will recognise their presence and stop until they are safely out of the way.Tree measurements system System to scan the orchard canopy and output t

106、ree measurement estimations.These measurements can then be used to estimate tree row volume for efficient spraying.The final system can reliably output tree height,tree width,and tree density.Is comprised of:1.2D(SICK LMS100)or a 3D(Velodyne VLP-16)2.Jetson Xavier(computer)3.Novatel SMART7 GPS recei

107、ver The point-clouds outputted by the LiDAR is processed in real-time by the on board computer utilising filters for identified Final report-AP16005 16 Hort Innovation tree netting as well as using the row-ends to accurately create a 3D map of the orchard.This data is then output in standard Raster

108、files that can be viewed and manipulated in any GIS software.The data can be used to calculate Tree Row Volume for that particular scan,giving the grower accurate insights into their orchard at different times throughout the flowering season.It has been demonstrated that LiDAR can be used to produce

109、 accurate canopy maps including tree height,tree width,canopy density,and detecting missing sections.It has also been shown that this can be achieved in real time in different orchard environments.The feasibility of eliminating the need for expensive hardware has also been explored with the use of c

110、amera algorithms instead of needing the LiDAR data.Doing so brings down the overall cost of the Collection Rig making it more attractive in price to more people.Flower stage estimation system Data collection system that could process images taken of an apple orchard and process in real-time to produ

111、ce data map viewable on a website.The final system developed for estimating the flower stage of apples trees has the following features:Capture full height images of every tree in an orchard.Process images in real-time travelling at 8 km/h.Get live-stream images and processed data to the Control Int

112、erface.Process multiple blocks in any order and any row-width.Instantly upload outputted data to the Cloud,ready to view on the Decision support tool.The tangible outputs consists of:Compact Collection rig that can fit on a Polaris with the following parts:Laser-cut frame Custom wiring harness Off-t

113、he-shelf waterproof case 2x Flir Blackfly S cameras with Fujinon 8mm(focal length)lens and CEI rugged enclosures 4x Phoxene strobe lights Novatel Smart7+Relay7-400 radio Teltonika 4G/Wifi router with John Deere antenna Jetson Xavie Carrier board Full Computer image that can be flashed onto AI Final

114、report-AP16005 17 Hort Innovation computer GitHub repository with core collection code Advanced AI image detection models trained on 3 years of data in Shepparton and Batlow An IP agreement has been developed to licence this technology to a commercialisation partner.See below for further details.Tes

115、ting pipeline for training datasets Quick turnaround process for collecting images with ground-truthed data and training a new model In conjunction with the development of the Collection Rig a quick turnaround training pipeline was developed to rapidly re-train the algorithm models as new data sets

116、were captured and ground-truthed during the season Online solution for viewing and creating spray maps Online database and website for storing and viewing produced flower stage/intensity maps from the collection Rig as well as outputting spray maps.This Decision support tool is for being able to vis

117、ualise the stage and intensity maps to the grower to help them visualise the variability in timing and flower load of their orchard.All the data collected during the 2021 season at the orchards in Batlow is available to view online through the decision support tool website.The source code for the we

118、bsite,data formats and database configuration will be provided in a GitHub repository in conjunction with the commercial partner taking on the flower stage software/hardware.The website was created to make the data more accessible and easier to visualise to help growers make more informed decisions

119、utilising their spray program with the data collected from the Collection Rig.IP agreement for technology transfer SwarmFarm and Hort Innovation have developed an IP licence to transfer the Apple Snapper technology developed in this project to a partner that will deliver it to industry.The licence h

120、as been developed to reduce the roadblocks that normally exist around IP transfer Commercialisation is currently being explored.At the time of writing this report,several conversations were underway with third parties,to ensure the transfer of the Apple Snapper technology and will be widely availabl

121、e to industry For commercial opportunities please reach out to Hort Innovation.Variable timing of application This is a concept that was developed throughout the project to target all areas of the orchard at the desired time rather than a one size fits all approach.A grower selects the growth stage

122、they want to target for primary thinning,e.g.60%king bloom.Any tree that is 60%KB or above would be targeted on this application while areas that have not reached the desired 60%KB would be targeted in Initially the project called for variable rate and variable concentration application to individua

123、l orchard trees.However,after talking to growers and working in the industry for two years,it was apparent that the biggest factor in fruit load management was the timing of the application.Not necessarily how much of a chemical was used,or the water volume.Also,variable concentrations rates would m

124、ean off-label use of chemicals which goes against APVMA regulations.SwarmFarm developed this concept after releasing the impact timing had on fruit load management.This concept was already in use by some farmers,who would manually switch the sprayer on/off as they drove through the orchard,relying o

125、n their experienced eye to make the call.The technology that has been delivered gives all growers including the more inexperienced operators the ability to variably time their thinning application.Over time this will reduce the variation of Final report-AP16005 18 Hort Innovation another application

126、.fruit throughout the orchard block.Final report-AP16005 19 Hort Innovation Outcomes Table 2.Outcome summary Outcome Alignment to fund outcome,strategy and KPI Description Evidence A product developed and ready for commercialisation Outcome(1)in project proposal The final product has the capabilitie

127、s to:Collect and process camera data at any stage of flowering Collect and process LiDAR scans to determine,height,width and density of canopy Using machine learning algorithms,determine the variation in crop load and the stage of growth(percentage of flowering)of each tree in the block Output commo

128、n GIS data formats Present a simple intuitive user interface to the grower/agronomist via a webpage Visualise tree height,width and canopy density,and allow growers/agronomists to calculate tree row volume,or similar spray calculations using these data layers Visualise the flower load and flowering

129、stage(percentage flowering)throughout an orchard block.Allow the user to develop a site-specific management plan(prescription spray maps)by selectively combining tree metrics,Project staff interacted with growers and agronomists from the industry regularly throughout the project.Feedback from grower

130、s made it evident that the original project calling for flower counting and variable concentration spraying was not actually what the industry needed.SwarmFarm took this on board and undertook a variation to the project scope to ensure the final product delivered was a solution the apple industry ne

131、eded.Final report-AP16005 20 Hort Innovation data on crop load and flowering stage Export a standard format prescription map that can be accepted by a variable rate spray system The product will provide functionality as specified above Outcome(2)in project proposal There are multiple technologies wo

132、rking together to deliver the final product target of tree by tree flower thinning.SwarmFarms role was to produce a system so anyone could use the algorithms developed by University of NSW to improve productivity on their orchard.The functionality has been simplified in response to industry demand.M

133、inutes after finishing a scan the prescription map can be loaded into the sprayer and the spray operation commenced.Timing is crucial in any agricultural enterprise and the feedback received was heavily focused on the technology being functional.Product features:Ruggedised and weatherproof modular u

134、nit Mounts to either a standard frame tractor,utility farm vehicle(i.e.Polaris)or SwarmFarm robotic platform On-board GNSS positioning system Onboard data storage,image and LiDAR processing Mobile data connection for data export Provide functionality within the system that will allow the grower/agro

135、nomist to create customisable management zones,within an orchard.Each management zone may consist of single or multiple trees,and can be selected based on logistical or hardware(sprayer)limitations Allow the user to identify heavier and lighter crop loads throughout an orchard block and develop a hi

136、gh-level management plan for the season to determine the spray strategy,i.e.more and/or earlier spray Final report-AP16005 21 Hort Innovation applications in heavier management zones Allow the user to then selectively filter for management zones that are at a particular stage of flowering(i.e.show a

137、ll trees greater than 50%flowering)and generate a prescription map from these zones Identify scouting locations(regions of high or low density)that may need further investigation The product is supported by 3 years of trial data Outcome(3)in project proposal Over the last three flowering seasons the

138、re has been 3 times weekly scanning with the data collection rigs.This data was then reviewed and used to start discussion with growers/agronomists.Each scan had a series of quads that were counted by professionals in the area.These were cross referenced to ensure the accuracy of our collection alon

139、g with the flower detection algorithm.Sprayer testing has been carried out over the last three years in and out of orchards to ensure the ability of the variable timing of application sprayer to accurately target individual trees.The technology has been rigorously tested in different locations,orcha

140、rds and tree varieties.The sheer volume of data collected and processed is proof to the industry of the quality of the data collection system and variable timing of application sprayer.A high-level of awareness of the product within the apple industry measured by adoption rate/volume Outcome(4)in pr

141、oject proposal Industry awareness has been increased through industry engagement at every step possible,however ROI and adoption will need to be reported on by Hort Innovation at a later stage.Videos and conversations have been the best way to gauge the interest and potential adoption by industry.In

142、 depth conversations with growers created the list of information that needed to be covered in the videos.Correctly conveying the ruggedness,simplicity and benefits of the full stack of technology from the data collection rig to the sprayer substantially increased the awareness in industry.Short vid

143、eos packed with technical information mean growers can see and understand how the system works and the benefits it will have to their operation.Sprayer setup and configuration Creating a build-guide for distribution among sprayer manufacturers and growers Conversations with many growers highlighted

144、a lack of understanding of what the currently available spray Final report-AP16005 22 Hort Innovation within the industry.The guide will cover installation,configuration,setup and operation of the key components of the sprayer system.The aim of the industry guide is to facilitate a larger uptake in

145、knowledge on how to utilise the large quantities of variability data available to growers.Giving this information freely to growers will increase the demand for data collection of different types within horticulture as a whole.technology is capable of doing with regard to variable rate application f

146、rom standard shapefile spray maps.Proving that tree by tree application is possible will raise awareness that this technology is here now and will create a demand for more data collection around orchards,not just flower stage and intensity.Sprayer hardware off the shelf and available The use of off-

147、the-shelf hardware gave confidence to growers that the technology and support was tried and tested and could be relied upon.Sourcing information from different industries such as broad acre in Australia and overseas,the latest technology was found and tested to see if it could apply to Horticulture.

148、Improved grower understanding of variability within their orchard Data collected from orchard blocks was regularly displayed to Growers in different formats to generate feedback on how best to display data in a user-friendly and effective way.Highly interactive data maps on the Decision support tool

149、 allow growers to see intensity and stage variability in their orchard with the click of a button.Feedback from growers and agronomists was by-annual bearing was the biggest issue.Therefore,statistics as a whole block or row was not of value.Being able to see and target overperforming trees and noti

150、ce underperforming trees in time will allow growers to have a much more consistent product to market year in year out.Grower confidence in technology Field walks were conducted with growers to show the data outputted by the Intensity and Stage Algorithms.Growers could see from tree to tree how accur

151、ate the produced maps were in comparison to the grower feel of their own orchards.Having growers in the loop providing feedback ensures the product will be suited to its market.Having growers looking at the tree in person while looking at its photo in the decision support tool built confidence in th

152、e technology.Working with agronomists meant growers in the area were represented through their technical representative,not just those who had hands-on experience with the technology.Final report-AP16005 23 Hort Innovation Recommendations SwarmFarm has compiled a list of recommendations for future w

153、ork that could be undertaken in this area to further benefit Apple and Pear growers.Remove the dependency of LiDAR on the canopy mapping system.SwarmFarm believes the Flower Stage Detection algorithms could be updated to include canopy mapping estimation.This would reduce the overall cost of the uni

154、t.Hort Innovation has funded an excellent R&D project and SwarmFarm has built and delivered the technology ready for commercialisation.We believe the next step of this project should be a fully funded ROI with a commercial partner of the complete system.A three year project that utilises this techno

155、logy and looks at thinning results,fruit yield,fruit quality and orchard variation would give growers confidence in the technology.We believe this is needed for large scale updates.Other companies collecting data using LiDARs,cameras,drones and satellites should be engaged to produce maps executable

156、 by spray systems such as the one used in this project.Continued research and development into precision agriculture technology suitable for orchard thinning.Conversations with Raven engineers in the USA spoke of new technology that could be more suitable for orchards.Newer spray control systems sho

157、uld also be looked into due to the fast moving nature of agriculture technology development,the sprayer control system used in this project was just an example of what can be done.Crowd sourced image collection with ground truth data is another way to expand the robustness of the algorithm in differ

158、ent growing regions.This would increase accuracy in different climates and factor in multi-year growing cycles Creating an iPhone app that can be used for:Collecting training data I.e.farmer could take a photo of a tree and input the stage distribution from manual counts Quick scan of trees Growers

159、could take a photo of a few trees in the orchard and get feedback on flowering stage Final report-AP16005 24 Hort Innovation Intellectual property Intellectual Property available for licencing This project has generated the following intellectual property available for licencing through Hort Innovat

160、ion.1.Apple snapper software o Collects and stored datasets on the fly for algorithm training o Interfaces with external hardware to capture images at the correct location and spacing o AI Algorithms for detecting flower stage and canopy density o Real time AI analysis 2.Decision support tool o Clou

161、d infrastructure for secure storage of datasets and online syncing of processed data o Decision support webtool for farmers to interact with flower stage data for better decision making o Prescription map creation for spraying/targeting individual trees.Other Intellectual Property generated in the p

162、roject The project has also generated some know how that has been documented into an easy to access guide.This industry guide(appendix 6)documents the learning made throughout the project and gives detailed information on how to configure an orchard sprayer to target individual trees.Final report-AP

163、16005 25 Hort Innovation Appendices Appendix 1-Sprayer Measurements and Results of testing in orchard A series of tests were carried out to measure the accuracy of the nozzle activation and shutoff of the Raven system within an Orchard.The tests were carried out at 4&6 km/h,targeting a tree defined

164、as 1m wide.The distance between trees was around 1.5 meters.Water sensitive paper turns to a dark purple colour anywhere a droplet hits the piece of paper,it is usually used to determine the droplet size of the chemical being applied however we have used it to determine exactly where the droplets we

165、re landing and therefore the measurable accuracy of the sprayer.Test 1 Water sensitive paper was placed(in order)-900,-450,0,250,700 mm from the centre of the first tree,then a tree was not sprayed,then water sensitivity paper was placed-700,-250 and 250 mm from the centre of the second tree.The res

166、ults are shown below.(a)Run 1 first tree (b)Run 1 second tree Figure 6:The first recorded test run In a perfect world we would see no droplets on the water sensitive paper greater than 500mm from the centre of the tree.We found that wind had a significant effect on the location where the droplets la

167、nded.We also found that any water sensitive paper that was behind some foliage or a branch had many less droplets which made it difficult to determine the accuracy of the nozzle activation/deactivation.Test 2 Water sensitive paper was placed-800,-400,0,400,700 and 1000 mm from the centre of the tree

168、 at 2.5m off the ground and 1m off the ground.Final report-AP16005 26 Hort Innovation (a)Run 2 (b)Run 2 bottom results (c)Run 2 top results Figure 7:The second test run These results confirm the accuracy of the system,with most of the droplets within 500mm from the centre of the tree.The nozzles cou

169、ld be seen to be switching at the correct location,however again the wind and the fan of the nozzles saw some of the paper outside the tree area have droplets land on them.There seemed to be a trend with the paper after the target having droplets on them,to try to combat this issue we increased the

170、distance between the GPS and nozzles in the Raven software.Test 3 In the third test run,the distance between the GPS to the nozzles in the Raven software was increased by 100mm.This was to try to combat a trend which we were noticing that was believed to be late nozzle activation/deactivation.This i

171、n hindsight made the results worse,however the results are below.Final report-AP16005 27 Hort Innovation (a)Run 3 (b)Run 3 bottom results (c)Run 3 top results Figure 8:The third test run Test 4 The fourth test was carried out at 6km/h,the distance from the GPS to the nozzles was also shortened by 10

172、0mm(from the original measured setting)to make the nozzles turn on earlier with respect to the mapped location.The water sensitive paper was placed-700,0 and 700 mm from the centre of the tree,with extra water sensitive paper 2.5m and 3.1m off the ground located at the centre of the tree.Final repor

173、t-AP16005 28 Hort Innovation Figure 9:The results from the fourth test We saw an improvement in these results however there was still a delay in the nozzle activation/deactivation.Therefore,the GPS to nozzle distance was increased in the Raven software.Test 5 The fifth test was also carried out at 6

174、km/h,the distance from the GPS to the nozzles was also shortened by 200mm(from the original measured setting)to make the nozzles turn on earlier with respect to the mapped location.The water sensitive paper was placed-700,500,700 and 900 mm from the centre of the tree,with extra water sensitive pape

175、r at the centre of the tree 3 meters from the ground.Figure 10:The results from the fifth test These results were better and to the naked eye it looked very accurate.However,the distance to the GPS from the nozzles in the Raven Software was shortened again.Final report-AP16005 29 Hort Innovation Tes

176、t 6 The sixth test was also carried out at 6km/h,the distance from the GPS to the nozzles was also shortened by 300mm(from the original measured setting)to make the nozzles turn on earlier with respect to the mapped location.The water sensitive paper was placed-700,-500,500,700 and 900 mm from the c

177、entre of the tree.Figure 11:The results from the sixth test This test showed the results we were looking for,with around 100mm accuracy to the mapped location which is what we saw with the fence line tests.Test 7 We were happy with the previous results and as there was a slight breeze,we decided to

178、try the next test from the opposite direction,previously we had been heading mostly south so for this test we were heading mostly north.The distance between the GPS and nozzles in the Raven software was left at 300mm shorter than the original setting.The water sensitive paper was placed 900,700,400,

179、-400,-700 mm and an extra sheet down very low 900mm from the centre of the tree.Final report-AP16005 30 Hort Innovation Figure 12:The results from the seventh test These results brought us back to over 200mm accuracy,this was due to the wind so for the next test we brought the distance back slightly

180、 and tested it going south again.Test 8 The eighth test was carried out at 6km/h and heading toward the south again.The distance between the GPS and the nozzles in the Raven software was shortened to 150mm shorter than the original setting.The water sensitive paper was placed-700,-400,400,700 and 90

181、0 from the centre of the tree.Final report-AP16005 31 Hort Innovation Figure 13:The results from the eighth test The results were as expected,the nozzle activation and deactivation was a little late,due to the wind.Test 9 The ninth test was carried out using the same settings,speed and water sensiti

182、ve paper setup.Final report-AP16005 32 Hort Innovation Figure 14:The results from the ninth test The results were also as expected,the nozzle activation and deactivation was a little late.This lead us to the conclusion that the results we saw in the fenceline test were comparable in the orchard.We c

183、ould have good coverage throughout the targeted tree with about 200mm of overlap in slightly windy conditions.Test 10 The final test was using the same settings,it was designed to test the consistency and accuracy of the sprayer in the toughest scenario.We made a prescription map that only sprayed e

184、very second tree on both sides.Six trees in a row had water sensitive paper pinned to leaves in the centre of the trees.Trees 2,3 and 4 also had a piece of paper put in the row behind them to see how much chemical would be going through the trees.The results are displayed below and are as expected w

185、ith some chemical going to the next row over through tree 4 with less foliage.Final report-AP16005 33 Hort Innovation Figure 15:The results from the tenth test 2.5 Commercialisation After the rigorous testing and feedback from growers and agronomists,it is clear there is a market for variable timing

186、 of application primary thinning in conjunction with the flower stage detection system.This season the sprayer design was focused on commercialisation,using common”off the shelf”components and documenting the build procedure.This build procedure has been shared with Silvan,to ensure feedback was acc

187、ounted for when building the sprayer used for this testing.The simplicity of the design and the industry guide will make it possible for anyone with some sprayer experience to update the plumbing and install the Raven control system.This guide also outlines the method for setting up the software par

188、ameters in the Raven display and how to load in a prescription map that is generated using the Decision Support Tool.The industry guide to integrate the Raven Hawkeye control system with the silvan airblast sprayer is attached in appendix.Final report-AP16005 34 Hort Innovation Appendix 2-Autonomy i

189、n an orchard This work package is developed in-kind by SwarmFarm Robotics.Figure 16:SwarmBot Papa running in the Orchard SwarmBot Papa was used to mow orchard blocks in Batlow NSW in 2021.Operating in 3.5m wide rows with a 2.45m Ben Wye slasher.The SwarmBot was customised to suit orchard conditions,

190、with an increased turning circle,smaller lug tyres for increased torque in the rough terrain of the Batlow region and fitted with a high precision GPS for improved position accuracy within an orchard.Data was also collected using the on-board obstacle detection cameras to help build out a viable sol

191、ution for robust safety in an orchard environment.Figure 17:GIS maps generated for the orchard at Batlow NSW Demonstration of supervised autonomous traversal of several orchard rows with SwarmBot SwarmBot was able to successfully traverse between orchard rows and different orchard blocks.Papa is lim

192、ited to a 2.7m turning radius,which means that a skip-row pattern was needed to be established so the machine could easily turn in/out of each orchard row.This is a common approach used by farmers when spraying/mowing the orchard.Final report-AP16005 35 Hort Innovation Demonstration of autonomous mo

193、wing of a block with SwarmBot.Report on cost/benefit analysis of autonomous mowing in orchard SwarmFarm has successfully demonstrated autonomous mowing within a few orchard blocks in Batlow NSW.Papa successfully operated in both 3m and 3.5 row spacings.After a few trials it appeared that the safest

194、operating speeds for an autonomous mower within an orchard is around 6km/hr.In a 3.5m row orchard,the robot is able to achieve around a 1.5ha/hr mowing rate.Autonomous machines can operate 24 hours a day,meaning 36 hectares a day could be completed.In some situations the grower will offset the mower

195、 to get closer to the tree bases as well as for different spaced blocks.In this situation the machine would have to drive the same path twice,offset to the left,then to the right.This method would half the number of total hectares done for the day.Having an autonomous robot mowing unsupervised frees

196、 up the grower to get on with other tasks around the orchard rather than being stuck in a cab all day.Final report-AP16005 36 Hort Innovation Appendix 3-Flower Stage Estimation The following are the upgrades that were undertaken during the 2021 year.The main focus was robustness and speed of the sof

197、tware as well as developing the full system pipeline for making the process of data collection to spraying the application map as smooth as possible.Figure 18:Example usage for the Apple Snapper Software The software for the collection rig was refactored to focus on speed and reliability.Software wa

198、s developed to more accurately trigger the cameras at exactly the right locations when driving down the rows without the need for row end points.This guaranteed that the location of every image captured and processed was correct to give the most up to date information to the Decision support tool.Us

199、er Interface The User interface was also refactored to make collecting data much easier for the growers.Live images along with the processed output were streamed to the web-app in real-time allowing the grower to see what the collection rig was capturing immediately.This gave confidence to the growe

200、rs that the values the algorithms were outputting were accurate and consistent.Upload to Cloud Both Collection rigs have the ability to automatically upload output maps at the completion of each scanned block.Standard Shapefiles and geoJSON file formats were used to store the collected data.The data

201、 was recorded to file in real-time as the images were processed,limiting the loss of data if the device lost power for any reason.Hardware This year the hardware side of the flower detection unit was improved,consolidated and finalised into a commercial ready and scalable product.The structural aspe

202、ct is made from laser cut and folded steel,all designed and modelled on CAD programs to produce workshop drawings.This means the components can be accurately and cheaply mass produced anywhere in Australia or Final report-AP16005 37 Hort Innovation across the world.The structural components have bee

203、n designed to bolt straight onto a stock standard Polaris Ranger,and with the modification of 2 parts this could be adapted to suit any side by side on the market.The overall design is quick to fabricate and install,embracing a modular design to quickly installed or removed by one person.Even when t

204、he flower detection unit is installed on the side by side,there is no compromise on everyday use,with the exception of slightly reduced cargo space.The whole unit is water and weatherproof and built to withstand the rigorous environment of the orchard.Wet,hot or dusty conditions or impact from branc

205、hes have no impact on the performance or longevity of the flower detection unit.A 3D CAD model of the final design The wiring circuit was made by Harness Master in Maitland NSW.It is documented in electrical schematic drawings suitable for mass production.Installation is simple and quick,only needin

206、g to hook up to the side by side battery and 2 switches to turn on the computer and the lights.This again means someone without any engineering or electronics experience could easily assemble the electrical components of the flower detection system.Structural Frame The unit had to be designed mainly

207、 around the position of the cameras.As the cameras are used to capture images of trees that are over 4m high in rows that are 3.5m wide it was important to have the cameras located as far from the tree they are capturing,and as close to the centre as possible.This meant the cameras had to be located

208、 around 2100mm from the ground,and as close to the centre of the data collection rig as possible.Once the ideal position of the cameras were found,the lighting setup was designed to be compact.The strobe lights were pointed at the top and bottom of the area being photographed,this created the bright

209、est,most consistent lighting across the tree.Lights The final selection was the Phoxene sx-3 strobe light.Strobe lights allowed us to maximise the limited electrical output(40 amps)of the side by side platform the system was designed for.The strobe lights are much brighter than common flood or spot

210、lights.This increases the quality of the lighting which allowed us to travel at greater speed and capture crisper images,and collect consistent data during the daytime.The strobe light only pulses for.00025 seconds each time the camera triggers an image.The factory camera software is used to trigger

211、 the lights which is a method widely used in machine vision.This means the simplicity of the system is maintained but the quality and variability of the environments it can be used in are vastly improved.Increasing the brightness of the lights meant this year the daytime data was much more consisten

212、t.In previous years,the lighting array was not bright enough to overpower the sun.This meant the output from the algorithm had some variability due to the overexposure of images looking into the sun.The lights are not waterproof so a waterproof enclosure was developed that included diffusers for the

213、 lights.This is used to reduce hotspots in the images(areas of overexposure where the clusters cannot be counted).Two lights are used for each side of the vehicle,one points at the tops of the tree and one pointing at the bottom of the trees which gave the brightest and most even lighting across the

214、 orchard blocks we worked in.Camera enclosures The cameras used were not ip 68 rated,therefore additional camera enclosures were used to protect the sensors and the lenses.The camera enclosures are much lighter and easier to remove/connect to the flower detection system due to the M12 connectors tha

215、n the orca enclosures used last year.The CEI Machine Vision Camera Enclosure 55m Series Round were brought in from the USA due to their compact design and price.The CEI camera enclosures used in 2021 Final System Components As reported in the previous sections a complete solution for producing flowe

216、r stage maps was developed that could:Capture full height images of every tree in an orchard.Final report-AP16005 38 Hort Innovation Process images in real-time travelling at 10km/h.Get live-stream images and processed data to the Control Interface.Process multiple blocks in any order and any row-wi

217、dth.Instantly upload outputted data to the Cloud,ready to view on the Decision support tool.The final parts list is as follows:Laser cutting and folding list:2x 50001 1x 50002 1x 50003 6x 51000 1x 51003 1x 51004 1x 51005,1x 51006 2x 51007 1x 51008 1x 51009 2x 51010 2x 51011 4x 51012 1x 1502 Ebox enc

218、losure:Pelican case BG057032032.Wiring harnesses:HM-SF-AS2-01 HM-SF-AS2-02 HM-SF-AS2-03 HM-SF-AS2-04 2x HM-SF-AS2-05 Strobe lights:4x Phoxene sx-3 Harsh.Computer:Jetson Xavier-NRU-120S.Final report-AP16005 39 Hort Innovation Cameras:2x FLIR Blackfly S(BFS-PGE-88S6C-C).Camera Lens:2x Fujinon FUJCF8ZA

219、-1S Camera Enclosures:2x CEI EN-38R-A1-82.LiDAR:Velodyne VLP-16.$6780 OR SICK LMS111-10100.GPS:Novatel Smart7+Relay7-400 radio.Relay PLC:IFM CR0431.4G/WiFi Router:Teltonika RUT240-LTE.4G/WiFi antenna:John Deere PFA10882.Final report-AP16005 40 Hort Innovation Appendix 4-Decision Support Tool Website

220、 Features Accessibility and time to interaction were the two goals that influenced how the final decision support tool would look.Considering this it was decided that a website was the best option as opposed to a QGIS extension or a Windows App.A website allows the data to be accessed by anyone with

221、 a device connected to the internet and offers are large amount of custom user experience and interactivity.The key features that the Decision Support Tool Website offers are as follows:Access to all collected data.Figure 21(1)Can toggle between Top,Bottom,or Average for each tree.Figure 21(2)Change

222、 between Intensity,Stage,and Spray Map visualisation.Figure 21(3)Dynamic colour scales depending on Intensity or Stage.Figure 21(4)See all previous data collected for current Block 21(5)Change the cutoff for percentage past the selected stage.Figure 21(6)Visualise average stage distribution for whol

223、e block on line chart.Figure 21(7)View all data collected on a map with Satellite and Normal maps.Figure 21(8)Figure 21 Final report-AP16005 41 Hort Innovation Figure 22 Insights Flower load on a tree is an important metric to a grower in determining the overall spray strategy.The collection rig ana

224、lyses the top and bottom sections of every tree and is able to detect white flowers,pink buds,and green tips and output an intensity value based on this data.These intensity maps can be used to quickly see which sections of the orchard block are producing a heavy load of flowers.This can be shown in

225、 Figure 22 where the areas of green and blue are much heavier than the areas in the red and yellow.One of the driving factors in determining when to spray an orchard block with chemical thinner,specifically ATS,is what flowering stage each bud is at.Our algorithms categorise the different stages int

226、o 8 categories;Green Tip,Half-inch Green,Tight Cluster,Pink Bud,Balloon,King Bloom,Full Bloom,and Petal Fall.This means that there is a key window of when the spray application applied is effective at a tree by tree level.For example an agronomist may count all buds on a tree and count 200 buds at o

227、r below Balloon stage,100 buds at King Bloom,and 100 at full bloom.This means that 50%of the buds on the tree are at or past the King bloom stage.As we are able to accurately assess the top and bottom of every tree and return a percentage value for each stage in each section,we can calculate this pe

228、rcentage pass King Bloom,or for any stage,for the whole block at a tree by tree level.From this data you can quickly see what flowering stage each tree is at across the whole block.This can be seen in Figure 21 where the colour scale is a percentage from 0(Red)to 100%(Blue)past and including King Bl

229、oom showing that there are large areas of the block past 50%King bloom(green and blue).Both these data visualisations can help the grower understand how heavy the individual trees are with flowering buds as well as how far progressed they are in their flowering stage.The more data you gather during

230、the early stages of flowering,the more insights you get in seeing how fast or slow the whole block or areas of the block are progressing.Using this data along with weather data allows the grower to make more informed spray decisions of when,where,and how much they need to apply for every tree in the

231、 block.Spray Map Output Being able to see flower intensity and stage information for an entire block at a tree by tree level is a great resource on its Final report-AP16005 42 Hort Innovation own in helping the grower make informed decisions in regards to how they treat their block.This project take

232、s it a step further and allows the grower to analyse and output a spray map that can be loaded into a compatible sprayer and actually treat each tree individually based on the data collected.The grower is able to use a slider as shown in Figure 25 to choose which sections(Top and Bottom of every tre

233、e)they want to spray or ignore.This slider filters out every section that falls below the select percentage of past King Bloom.In other words,in Figure 25 the grower has select every section that has at least 70%of its buds at or past King Bloom,and wants to apply a water rate of 800L/ha for the top

234、 and 1200L/ha for the bottom.This level of application means the grower does not have to change water rates or turn on and off sections manually based on the maps,but can rely on the sprayer software to turn on and off at the exact position it needs.The Sprayer Software is able to use the same GPS a

235、s the Collection Rig to achieve this accuracy allowing for very precise application.Final report-AP16005 43 Hort Innovation Appendix 6-Airblast Sprayer Industry Guide Final report-AP16005 44 Hort Innovation Final report-AP16005 45 Hort Innovation Final report-AP16005 46 Hort Innovation Final report-

236、AP16005 47 Hort Innovation Final report-AP16005 48 Hort Innovation Final report-AP16005 49 Hort Innovation Final report-AP16005 50 Hort Innovation Final report-AP16005 51 Hort Innovation Final report-AP16005 52 Hort Innovation Final report-AP16005 53 Hort Innovation Final report-AP16005 54 Hort Inno

237、vation Final report-AP16005 55 Hort Innovation Final report-AP16005 56 Hort Innovation Final report-AP16005 57 Hort Innovation Final report-AP16005 58 Hort Innovation Final report-AP16005 59 Hort Innovation Final report-AP16005 60 Hort Innovation Final report-AP16005 61 Hort Innovation Final report-AP16005 62 Hort Innovation Final report-AP16005 63 Hort Innovation Final report-AP16005 64 Hort Innovation Final report-AP16005 65 Hort Innovation Final report-AP16005 66 Hort Innovation Final report-AP16005 67 Hort Innovation Final report-AP16005 68 Hort Innovation