Vineyard Computing:Sensor Networks inAgricultural Production
Mobile and pervasive computingtechnologies provide us withsome of the first opportunitiesto explore computing outsideclimate-controlled building en-
vironments. With this freedom comes an endlessvariety of environments that the research com-munity has just begun to explore as potential sitesfor technology use. The original pervasive com-puting systems used office spaces and officemobility as a jumping-off point for concept explo-rations.1 We pursued a different approach bylooking at work environments outside the office,including medical clinics, manufacturing plants,and farms.
This article discusses an extended study of vine-yard workers and their work practices to assess
the potential for sensor net-work systems to aid work inthis environment. The study’slarger purpose is to find newdirections and new topics thatpervasive computing and sen-
sor networks might address in designing tech-nologies to support a broader range of users andactivities. We expect that much of what weuncovered in this research will be useful to tech-nology design for outdoor environments, othertypes of agriculture, and mobile work environ-ments in general.
Previous research on sensor network applica-tions has frequently focused on partnershipsbetween technologists providing the sensor net-works and biological and environmental re-searchers studying habitats and endangered
species.2–4 As a potential user group, agricultur-alists are distinct from scientists doing habitatresearch. They focus on production rather thanexploratory research, so they’re not interested inspending time interpreting data. They want datathat recommends a course of action, somethingthat will save them time rather than create addi-tional work. Also, agriculturalists aren’t work-ing in remote or fragile environments. They inter-act closely and physically with crops, touchingand examining them each day. They know theycan’t farm remotely.
These two primary differences in work activi-ties and priorities between agriculturalists andbiologists indicate why our study is important inthe discussion of sensor network applications.The sensor network application requirements forbiological researchers aren’t the same as those foragriculturalists and others working on vineyards,farms, or other sites of agricultural production.
In addition to looking at a new category ofusers, our study is also distinguished by ourhuman-centered research approach. We usedethnographic methods including interviews, sitetours, and observational work to broadly under-stand the work activities and priorities of the var-ious roles working in a vineyard. This rigorousand holistic approach to what software devel-opers might describe as requirements gatheringwas particularly important because we werestudying a population with work activities verydifferent from our own. In contrast to previoussensor network implementation projects, our tar-get users weren’t researchers, nor were theyapproaching their work from a research per-
Using ethnographic research methods, the authors studied the structureof the needs and priorities of people working in a vineyard to gain abetter understanding of the potential for sensor networks in agriculture.
Jenna Burrell, Tim Brooke, andRichard BeckwithIntel Research
spective. Their mindset was one of pro-duction and optimization.
Ethnographic researchmethods
Our group uses a general researchapproach that focuses on studying peo-ple and practices before technologyinterventions are designed and put intoplace. We employ ethnographic methodsas a way to gather rich data about thepeople who inhabit environments thataren’t well understood as sites for tech-nology use. In this particular study ofvineyards, we looked at people’s rolesacross the entire value chain of wine pro-duction, with the belief that each rolerepresents a different relationship withthe vineyard and winery and differentinformation and interaction needs.
We conducted semistructured inter-views with vineyard managers, vineyardowners, winemakers, vineyard market-ing people, and wine sellers. We alsoconducted site tours and photographedvineyards, wineries, and wine shopsguided by our interview subjects. Dur-ing the busy season, some members ofour team became participant observersby joining work parties to help out dur-ing harvest and to put up nets to protectthe grapes from migratory birds.
After studying the vineyard as a poten-tial site for technology use, we movedinto a second phase of the project todevelop technology concepts and imple-ment a working sensor network. We cre-ated a series of interface designs and tech-nology-interaction concepts that wouldfit into this work environment on thebasis of our analysis of observations andinterviews. These concepts were pre-sented to vineyard managers, wine mak-ers, and agricultural researchers for fur-ther refinement and development, al-though they were not deployed as oper-ational user interfaces. A second phaseof research involved the limited deploy-ment of a working sensor network in a
local Oregon vineyard. The trial instal-lation involved the deployment of ap-proximately 18 motes for a period ofseveral weeks during the late summer of2002. This installation let us come faceto face with the challenges of installingcomputing technology and working withsensor networks outdoors. A third phaseof research, not described in this article,involved a much more ambitious sensornetwork deployment at a vineyard site inBritish Columbia.
Ethnographic research has proven inthe past to be a particularly successfulway of inspiring innovative technologyconcepts that directly address users’needs.5,6 We also found in our study ofvineyards that understanding the poten-tial users’ needs and work activities canprovide feedback on how existing sen-sor network hardware and software andother pervasive computing technologiesshould be configured and redesigned.Our primary goal is to uncover the impli-cations for sensor network design andresearch arising from user needs and thestructure of work activities in agricul-tural-production environments.
All this new digital dataPervasive computing technologies—
such as sensor network systems—give usnew capabilities for sensing and gather-ing data about an environment and newways to manage this data digitally. Wecan gain information about temperature,lighting levels, humidity, the movementand presence of people, and many otheraspects of the environment. However,these capabilities pose several questions
in the application space. What datashould we gather and how often? Whatlevel of computational interpretationshould we apply to the data? Howshould we present data to the user?When should the system act on data andwhen should action be left up to theuser? Our interviews and site visits gaveus concrete examples of the kinds of sen-sor network applications that would beappropriate and beneficial in an agri-cultural environment.
A combination of three factors pro-vided some answers to these questions:equipment capabilities, environmentalconditions, and user needs. Equipmentcapabilities include battery-life limits,processor power, types of available sen-sors, memory space, sensor accuracy,and radio frequency (RF) transmissionrange. These factors can make certainpotentially useful applications realisti-cally impossible. For example, someresearchers have described GPS local-ization as too power hungry to be real-istic in a sensor network. The environ-ment itself also provides answers toquestions about data gathering by pro-viding variation within a finite rangealong certain measurable axes.
In our implementation work, we dis-covered great variability across the vine-yard during the daytime but less varia-tion at night. For this reason, sensorreadings (a function that consumes a sig-nificant portion of the battery power)could be taken less frequently duringnight hours. Similarly, we discoveredthat variability of conditions across avineyard might be of greater concern
JANUARY–MARCH 2004 PERVASIVEcomputing 39
What data should we gather and how often?
What level of computational interpretation
should we apply to the data?
How should we present data to the user?
than RF transmission range in deter-mining the density of sensor placementin a network.7
Although many environments have aseemingly infinite variety of measurablecharacteristics, user needs provide anotherlimit on what actually should be measuredand how often. For example, we learnedfrom interviews that during the winter,there’s a risk of frost damage to vines, sothe vineyard needs a system to gather fre-quent temperature readings at night and
alert the manager when the temperatureis low. However, this need is seasonal, sofrequent night temperature readings com-bined with a real-time alert system are onlynecessary in the winter.
Once we know what data will be use-ful and relevant, the question becomeswhat kind of computational interpreta-tion do we need and what should we doonce the data is interpreted. At one endof the spectrum, the data can simply bedelivered raw. This approach has someobvious shortcomings for vineyardworkers: raw data might not suggest anycourse of action, or it might require sig-nificant effort to draw useful conclu-sions. In a production environment, thisextra interpretive work can be a signifi-cant time burden. At the other end of thespectrum, we might be able to thor-oughly interpret the data and performan action on the user’s behalf. Proactivecomputing recommends this approachto remove the user from direct interac-tion with the system.8 The benefit ofbeing proactive is that users aren’t over-burdened by system demands becausethey don’t interact with it directly.
In our interviews and site visits, vine-yard managers indicated what level ofdata interpretation would provide valueto them. These findings illuminate somecharacteristics of circumstances that rec-ommend proactive computing versus thealternative, which is providing inter-preted information without completingany sort of action. In any case, the datamust be actionable, a term used repeat-edly by one of the vineyard managers weinterviewed. He wanted the data to sug-
gest a tangible next step, so in our inter-face design work, we explored severalforms of actionable data.
The first was a map of powdery-mildew risk that could be calculated fromtemperature data readings gatheredthroughout the vineyard over a period oftime. A map generated in this way couldeasily demonstrate what areas of thevineyard were at the highest risk for pow-dery mildew and would let the vineyardmanager spray pesticides on the specificat-risk area to avoid problems. Unana-lyzed temperature data would have beeninsufficient for this purpose because youcalculate powdery-mildew risk using oneof a number of complex models that taketemperature data gathered over time asinput. Temperature data could also beused to make heat unit calculations thatvineyard managers use to get a sense ofthe grapes’ ripeness, which is a factor indeciding when to harvest.
Proactive computingProactive computing would suggest
that we design systems that interpretactionable data and then automatically
act on it. Examples in this study’s con-text are
• A vineyard equipped to spray itself inthe appropriate area when there’s arisk of powdery mildew
• An irrigation system that optimallyrations limited ground water
• An automated call to the workers tocome in and pick the grapes whenthey’re ripe
In fact, vineyards in New Zealand andAustralia mechanically (although notautomatically) harvest their grapesbecause there’s no labor pool to drawworkers from. In the US, grape-pickingteams primarily made up of migrantlaborers make more sense economicallyfor a farmer than investing in harvestingequipment.
However, automating the decision toharvest would be less than ideal, mainlybecause this is often a subjective andsocial decision based on incompleteinformation. This is precisely the kind ofproblem that humans are quite skilled atsolving. The winemaker plays the pri-mary role in deciding when to harvestand bases these decisions on the kind ofwine the vineyard intends to create. Avineyard manager plays a role in the har-vest decision by monitoring weatherreports for the threat of rain. Rain canruin ripe grapes by diluting the potentflavor of each grape or even causingthem to burst. If rain is in the near-termforecast, it will often lead to picking thegrapes before they’re perfectly ripe.
Because weather is so unpredictable,the decision to harvest is always a judg-ment call. Because many vineyards arelocated in the same area, there is also thechallenge of scheduling the local crew ofworkers to harvest the plants because allproximate vineyards typically decide topick at around the same time. There’ssocial pressure and competitivenessamong local vineyards. We talked to one
40 PERVASIVEcomputing www.computer.org/pervasive
S E N S O R A N D A C T U AT O R N E T W O R K S
Once we know what data will be useful and
relevant, the question becomes what kind of
computational interpretation do we need and
what should we do once the data is interpreted.
manager of several vineyards who usedthis to his advantage. To influence thedecision about when to harvest, he’dmention to the winemaker or vineyardowner that another vineyard had de-cided to harvest—thus pressuring thewinemaker or owner to follow suit. Itwas a subtle form of manipulation onthe manager’s part. So, the decision toharvest isn’t well suited to a proactive-computing approach because it resultsfrom social factors and incomplete dataand is not too difficult for humans to doon their own.
In contrast, other vineyard processesdo lend themselves to a proactive-com-puting model. For example, irrigation isa major issue in many vineyards. Anideal proactive system would optimizewater needs in different areas of the vine-yard with available water—particularlybecause water is a limited, sharedresource. Being able to water plantsmore selectively and precisely on thebasis of individual plant needs and avail-able water would save water. This typeof precision would be time-consumingfor a vineyard manager or worker, so aproactive system that does it on the man-ager’s behalf makes sense.
Similarly, dealing with pests is anotheropportunity for proactive computing. Itwouldn’t make sense, for example, todetect the presence of birds and alert themanager about the problem. This couldhappen many, many times throughoutthe day, and birds require a more imme-diate reaction than the manager can pro-vide. It only takes a minute or two for aflock of birds to do serious damage to agrape crop. A proactive approach woulddetect and respond to the bird presence,perhaps by shooting off a loud cannon.We were told in interviews that shoot-ing off a loud cannon periodically is oneapproach to dealing with bird threats.However, birds often get accustomed tothe same loud sound and continue to eatgrapes in spite of the cannons.
In these examples, proactive comput-ing plays an important role in dealingwith problems with two characteristics:those that require more immediate reac-tion than human capabilities can fulfilland those that require time-consumingactivities that would overburden vine-yard workers. In the case of irrigation,there’s a sophisticated level of optimiza-tion and computational work involvedthat computing power can help address.The financial investment involved in
equipping a vineyard with proactive sys-tems will be an important consideration.Some work that is repetitive and time-consuming, such as pruning, will con-tinue to be done by workers because,compared to the cost of labor, equipmentto do the task is too expensive or toocomplex to automate.
Our findings about the need foractionable data also led us to concludethat pervasive computing systems wouldneed to be designed with domain ex-perts’ involvement. For example, themodels one might use to illustrate pow-dery-mildew risks in our interfaces weredeveloped at the University of Califor-nia at Davis as part of the viticultureresearch program.9 It will most likely beagricultural researchers who take thecapabilities provided by ubiquitous com-puting technologies and connect themwith applied uses in the vineyard. Sen-sor-net equipment will also play a rolein domain-specific research by enablingresearchers to gather new data that couldlead to new knowledge about growinggrapes and other types of crops. Perhapsnot surprisingly, the researchers we’ve
been in contact with have already showngreat interest in using these technologiesfor research purposes.
Human touchpointsThe concept of human touchpoints
can be a useful way to think about userinteraction with pervasive computingsystems. We define a human touchpointas a portal that connects an individualwith the underlying system infrastruc-ture—in this case a sensor network—
either by supplying representations ofdata gathered by the infrastructure or byplacing the individual in the role of pro-viding input. What is characteristic ofpervasive computing systems is that asingle system can have multiple humantouchpoints of various types. In ourstudy of people in the vineyard and wine-making industry, we found that provid-ing a variety of human touchpoints wasimportant to address the different rolesand responsibilities of a heterogeneouspopulation of potential users thatincluded vineyard managers, hired tem-porary labor, winemakers, and vineyardowners.
How should data be presented to theuser? In what ways can users input datainto the system? In our interviews, weuncovered divergent sets of priorities andtasks associated with different roles. Thevineyard manager is an agriculturalistwho knows about pests, irrigation needs,and all the information associated withsuccessfully growing high-quality grapes.The manager also does business and per-sonnel management work and handlestime cards, budgets, and work delega-
JANUARY–MARCH 2004 PERVASIVEcomputing 41
Our findings about the need for actionable data
led us to conclude that pervasive computing
systems would need to be designed with
domain experts’ involvement.
tion. The winemaker is an artisan as wellas a scientist who uses both chemistryand good taste to transform grapes intofine wine and blend various wines intosomething interesting, complex, or mar-ketable. Vineyard workers in our area ofthe country are migrant workers andoften speak only Spanish. They work inteams during the harvest and are paidaccording to how much they pick. Theweight of picked grapes is tracked andassociated with each worker. From theseexamples, it’s apparent that
• The vineyard manager has manage-ment responsibilities that the wine-maker and vineyard workers do not.
• Winemakers have a subjective elementin their work process that vineyardmanagers and vineyard workers do not.
• Vineyard workers have a significantmanual-labor component in theiractivities and often don’t even speakthe same language as the vineyardmanager and winemakers.
Our interest in the roles engaging incollaborative work suggests the rele-vance of research in the domain ofcomputer-supported cooperative work(CSCW). Pervasive computing systems
often have some of the same character-istics as networked groupware applica-tions that CSCW researchers study. Bothfields must address the needs of hetero-geneous user populations working col-laboratively. However, pervasive com-puting systems differ from traditionalCSCW applications and technologiesbecause of their strong tie to physicalenvironments and physical activities thatinvolve and emphasize tool use and thelocation of activities, rather than infor-mation management and knowledgework as is typical in office environments.The needs of different roles in the vine-yard go far beyond providing access todifferent kinds of information; theseroles represent completely different workparadigms. Human touchpoints in per-vasive computing systems must negotiatebetween these paradigms.
For example, an interface that negoti-ates between these roles would providemultiple interdependent interfaces suitedto each role; it might address the vine-yard manager’s job of managing andcoordinating activities and paperwork.This task falls outside the weather- andenvironment-monitoring capabilities wegenerally assume sensor network sys-tems are good for. Through interviews,
we discovered that vineyard managersare interested in ways that technologycan help them with business manage-ment tasks, which often involve a lot oftime-consuming data entry. A sensor net-work could support management needsby tracking activities, personnel, andequipment through the vineyard andincorporating this data automaticallyinto budgets and time cards.
For the system to work, it wouldrequire multiple human touchpoints.One touchpoint would allow the man-ager to call up vineyard activity data andview it. A second touchpoint wouldallow vineyard workers to enter inputabout their activities into the system.Because the workers are primarily man-ual laborers, a system requiring them totype or explicitly enter data would inter-fere with their primary work activities.To resolve this issue, we developed theconcept of tagged tools as a way to helpgather data for budgeting and activitytracking.
For example, the manager might wantto know when and where the vineyardwas sprayed with pesticides to assess therisk of a powdery-mildew outbreak (seeFigure 1). If we instrument the vineyardwith a static sensor network, a pesticidesprayer tagged with an RF identificationtag or sensor network mote could beoperated by a vineyard worker andtracked as it sprayed areas of the vine-yard. The pesticide sprayer movingthrough the vineyard would then be theworker’s human touchpoint to serve asthe input device into the sensor networksystem. This input device would operatewithin the vineyard worker’s work par-adigm while still providing for the vine-yard manager’s information needs.
Similarly, pruning shears, shovels, andpicking boxes could also be given uniqueRF identification tags to track the loca-
42 PERVASIVEcomputing www.computer.org/pervasive
S E N S O R A N D A C T U AT O R N E T W O R K S
Figure 1. A vineyard manager’s interfaceshows a map of grapes in the vineyard,patches with high powdery-mildew risk,and areas that have been sprayed withpesticides via tractor.
tion and type of activity. By selecting andusing these tools, the vineyard workerwould provide the necessary input intothe system naturally and effortlessly. Theconcept of tracking workers’ movementthrough space has already been sug-gested as a useful tool for generatingbilling reports and time studies in custo-dial environments such as hospitals.10
We envision an instantiation of this ideawith the added concept of tagged tools toprovide an indication of workers’ activ-ities. The manager’s need to track activ-ities in the vineyard also suggests thatfocusing attention on developing local-ization algorithms for sensor networks—specifically tracking the location oftagged objects moving through a sensornetwork—is a research direction poten-tially useful for agricultural applications.
System architectureOur study also suggested different
types of interfaces that could be seam-lessly incorporated into the vineyard,including the tagged tools described ear-lier. However, our understanding of theworkflow also suggested some ways thatthe system infrastructure itself could bereorganized to optimize power manage-ment and equipment costs. Our efforts tocreate a working sensor network imple-mentation in a local vineyard gave ussome insight into the interplay betweenpower management, equipment costs,system architecture, and user needs.
Power management is one of the pri-mary issues in the design of sensor net-work systems intended to operate wire-lessly.11 An ideal system would be asensor network made up of devices thathave an extremely long battery life and
are automatically rechargeable or aretiny, disposable, inexpensive, and easilyreplaced. The concepts of Smart Dustand Paintable Computers are two pro-posals of this ideal vision.12,13 Becausewe believe that sensor networks are use-ful in the near term, we must realisti-cally face power management issues toavoid the worst-case scenario where bat-teries must be frequently replaced inhundreds or thousands of individualdevices. We have uncovered opportuni-ties for a systemwide approach to powermanagement by designing the softwareand system architecture to optimizepower management. However, ourmodest gains could be greatly improvedif the hardware were redesigned withthese systemwide configurations inmind.
Self-organizing ad hoc sensor net-works are generally considered thedefault system architecture, in partbecause they present more interestingcomputational problems for computerscientists to tackle. However, this archi-tecture assumes RF connections, oftenusing TDMA (time division multipleaccess, a technology for delivering digi-tal wireless service) between each moteand its neighbors. This arrangement ofsystem components requires enough
equipment to cover a space with a fullyconnected network. It’s an optimal archi-tecture for some types of applicationsbut is by no means the only one oralways the ideal arrangement of the net-work. Specifically, the self-organizingmultihop architecture that forwards datais the only architecture that makes muchsense for sensor network applications inremote, inaccessible environments.
We discovered that other system archi-tectures could be employed in vineyardsbecause they are neither remote nor inac-cessible. For example, one architectureused data mules to collect and transportdata from sensor network motes dis-tributed throughout the vineyard (seeFigure 2).14 From our interviews andobservations, we learned that during thegrowing season, workers move up anddown the rows a lot. In one vineyard,two family dogs also spent a lot of timegoing up and down the rows. Any ofthese moving bodies (even the dogs)could serve as a “data mule” by carry-ing a small device that simply and invis-ibly gathers data wirelessly from the sta-tic, distributed motes.
The data would be transmitted fromthe static mote to the data mule motewhenever the two motes are in physicalproximity and there’s new data to trans-
JANUARY–MARCH 2004 PERVASIVEcomputing 43
Figure 2. Data mule system architecturein the vineyard. (a) The motes recordenvironmental data and vineyard activities. (b) In the course of daily activities, the worker collects more dataonto the shovel. (c) The worker takes thetool back to the shed. (d) Back in the toolshed, the shovels upload their data to thecentral database.
(a)
(b)
(d) (c)
mit. This configuration does not repre-sent a distinct advantage in terms ofpower savings because static motes mustremain in RF listening mode in order tocommunicate with the data mule mote.However, it does save on equipmentcosts because we can distribute motessparsely throughout the vineyard, as theydon’t need to communicate with neigh-boring motes. Several applications in thevineyard lend themselves to in-networkdata storage and processing. When com-
bined with infrequent synchronization,this configuration has the potential forsignificant power savings and isamenable to a data mule solution. In par-ticular, in-mote distributed processingcould be used to calculate heat units thatdetermine the appropriate time to har-vest grapes. Because vineyard managersdon’t need to calculate heat units imme-diately—a latency of a few hours or aday is suitable—this application doesn’tneed a connected live-data sensor net-work. The need for application-specificdata aggregation and in-network pro-cessing is a unique requirement of sen-sor networks. This requirement distin-guishes sensor networks from traditionalwireless networks.15 We employed thisstrategy by using in-network data pro-cessing to reduce the quantity and fre-quency of RF transmission. To calculateheat units, we simply needed the dailyhigh and low temperatures. Each motegathered data once every 60 seconds andthen compared each data reading tostored high and low temperature pointsfor the day. A new low or new highwould replace the old one.
The only data that needed to be trans-mitted via RF was the absolute maxi-mum and minimum temperature for theday, because this is all that was requiredto calculate heat units. It should be notedthat we used Eeprom (electrically eras-able programmable read-only memory)in our implementation to store datalocally on the mote. To effect power sav-ings, a more power-efficient technology,such as flash, would be necessary. In fact,flash was built into the sensor network
motes we used, but writing data to flashwas not yet implemented in the TinyOSversion we were using. Our ability todesign a system that limited RF trans-mission and took advantage of in-net-work data processing rested on ourunderstanding of vineyard work. Welearned from talking to vineyard man-agers that heat-unit calculations wereuseful, actionable data that wouldimpact harvest. And we learned whatdata was required to make these calcu-lations. We determined that in-networkprocessing was possible because the sit-uation required only simple calculations.
Observing the constant movement ofpeople and dogs in the vineyard led us toconsider a system architecture thatrelied on data mules to reduce equip-ment costs. Vineyard managers’ use ofheat units to make harvest decisions ledus to use power-efficient in-networkdata processing. In these examples, thevineyard work patterns directly influ-enced our ability to create a useful sen-sor network application and to optimizeit to conserve power and save on equip-ment costs.
While exploring the poten-tial for sensor networks inagriculture, we gained anunderstanding of the
structure of vineyard work, the needsand priorities of the people who workthere, and the interaction between var-ious stakeholders and roles involved. Wefound that the way work is done in avineyard has direct implications fordesigning and configuring these envi-ronments’ sensor networks. Lookingtoward the future of sensor networkresearch, we can recommend severalareas where pervasive technology andsensor network researchers might focustheir efforts to address the needs and pri-orities of people working in agriculturalenvironments. One area of need is sup-porting alternative system architectures.
Because agricultural work involvesdaily movement through the farm, usingdata mules is a sensible approach toreduce equipment cost. We also needgood localization algorithms to trackequipment and people moving throughthe space. This capability would provideuseful data for management needs,including budgets, time cards, and gov-ernment-regulation paperwork. Agricul-tural environments also could use proac-tive-computing approaches that can acton the user’s behalf for applicationsrequiring a faster-than-human responsetime or that require precise, time-con-suming optimization. Research on opti-mized networks to loop sensor data withactuators would provide for proactiveapplications. Irrigation, frost detection,and pest detection are all examples ofapplications in agriculture that wouldbenefit from proactive approaches.
This article has not described a single,comprehensive solution for equippingagricultural environments but a varietyof sensor network configurations andapplications that can address differentpriorities in the vineyard. Some of thesensor network configurations and fea-
44 PERVASIVEcomputing www.computer.org/pervasive
S E N S O R A N D A C T U AT O R N E T W O R K S
The way work is done in a vineyard
has direct implications for designing
and configuring these environments’
sensor networks.
tures we’ve described are compatiblewith each other and some aren’t. Forexample, a sparse distribution of sensornetwork motes using data mules for dataforwarding won’t support localizationalgorithms that rely on triangulation.Different system configurations will varyby cost and capabilities. In practice, therewill likely be a plurality of useful sensornetwork systems employed in agricul-tural environments to address differentpriorities.
For example, a simple sparse networkemploying data mules might be a useful,inexpensive entry-level system that canbe upgraded later to include more motesand provide precise localization capabil-ities. Similarly, a data mule system archi-tecture will not support proactive com-puting applications that require real-timeresponse. However, agricultural workdepends on seasons and time of day, so asensor network that can self-configureaccording to temporal factors could com-bine some of these approaches. Forexample, a proactive system could mon-itor for frost during winter nights or forbirds during bird migration. Other timesof year, the system would use a data muleapproach. These examples suggest thepotential for several creative strategiesfor combining capabilities and systemconfigurations.
Taking a high-level view, the interfacedesign and implementation of humantouchpoints in the sensor network infra-structure must take into account collab-orative work environments and providemediation between vineyard managers,owners, workers, and winemakers.Research in any of these areas will be use-ful in the eventual development of sen-sor network technologies as consumerproducts for agricultural monitoring.
REFERENCES1. R. Want et al., The Parctab Ubiquitous
2. A. Cerpa et al., “Habitat Monitoring:Application Driver for Wireless Communi-cations Technology,” Proc. ACM SIG-COMM Workshop Data Comm. in LatinAmerica and the Caribbean, ACM Press,2001, pp. 20–41.
3. A. Mainwaring et al., “Wireless Sensor Net-works for Habitat Monitoring,” Proc.ACM Int’l Workshop Wireless Sensor Net-works and Applications, ACM Press, 2002,pp. 88–97.
4. E. Biagioni and K.W. Bridges, “The Appli-cation of Remote Sensor Technology toAssist the Recovery of Rare and Endan-gered Species,” Int’l J. High PerformanceComputing Applications, vol. 16, no. 3,Aug. 2002, pp. 315–324.
5. T. Salvador, G. Bell, and K. Anderson,“Design Ethnography,” Design Manage-ment J., vol. 10, no. 4, Fall 1999, pp. 35–41.
6. S. Lewis et al., “Ethnographic Data for Prod-uct Development: A Collaborative Process,”Interactions, vol. 3, no. 6, Nov./Dec. 1996,pp 52–69.
7. D. Estrin et al., “Instrumenting the Worldwith Wireless Sensor Networks,” Proc. Int’lConf. Acoustics, Speech, and SignalProcessng (ICASSP 2001), vol. 4, IEEEPress, 2001, pp. 2033–2036.
8. D. Tennenhouse, “Proactive Computing,”Comm. ACM, vol. 43, no. 5, May 2000, pp43–50.
9. “Crop: Grape Disease: Powdery Mildew,”Disease Model Database, 2001, www.ipm.ucdavis.edu/DISEASE/DATABASE/grapepowderymildew.html.
10. G.T. Carroll et al., Time and AccountingSystem, US patent 4549264, Patent andTrademark Office, 1983.
11. J. Lifton, “Pushpin Computing SystemOverview: A Platform for Distributed,Embedded, Ubiquitous Sensor Networks,”Proc. 1st Int’l Pervasive Computing Conf.2002, LNCS 2414, Springer-Verlag, 2002,pp. 139–151.
12. W. Butera, Programming a Paintable Com-puter, doctoral dissertation, MIT MediaLaboratory, 2002.
13. D. Culler et al., “A Network-CentricApproach to Embedded Software for TinyDevices,” Proc. 1st ACM Int’l Conf. Em-bedded Software (EMSOFT 01), ACMPress, 2001, pp. 114–130.
14. R. Shah et al., “Data MULEs: Modeling aThree-Tier Architecture for Sparse SensorNetworks,” tech. report IRS-TR-03-001,Intel Research, 2003; www.intel-research.net/Publications/Seattle/012220031206_114.pdf.
15. D. Estrin, “Next Century Challenges: Scal-able Coordination on Sensor Networks,”Proc. 5th Ann. Int’l Conf. Mobile Com-puting and Networking (MobiCom 99),ACM Press, 1999, pp. 263–270.
For more information on this or any other comput-ing topic, please visit our Digital Library at www.computer.org/publications/dlib.
JANUARY–MARCH 2004 PERVASIVEcomputing 45
the AUTHORS
Jenna Burrell is a doctoralstudent in sociology at theLondon School of Econom-ics. As an intern and later asa full-time employee withIntel’s People and PracticesResearch Group, she didresearch on animators, col-
lege students, fab technicians, and vineyardworkers. She received a BA in computer sciencefrom Cornell University. Contact her at 2111NE 25th Ave. (JF3-377), Hillsboro, OR 97124-5961; [email protected].
Tim Brooke is an interac-tion designer with the Peo-ple and Practices ResearchGroup at Intel. His currentresearch focuses on the kindof user experiences thatubiquitous computing canenable. He received an MA
in computer-related design from the Royal Col-lege of Art in the UK and is a member of theInstitute of Electrical Engineers (UK). Contacthim at 2111 NE 25th Ave. (JF3-377), Hillsboro,OR 97124-5961; [email protected].
Richard Beckwith is aresearch psychologist withIntel’s Corporate TechnologyGroup. He works in the Peo-ple and Practices Researchgroup, whose main focus isto find new uses for technol-ogy. He received a PhD in
developmental psychology from Teachers Col-lege, Columbia University. Contact him at 2111NE 25th Ave. (JF3-377), Hillsboro, OR 97124-5961; [email protected].
