Приложение 2

к пис. № 1015/04-8 от 28.07.2014г.

The Poseidon Project: An Introduction

The use of Cognitive Computing for real-world societal challenges

Robert-Jan Sipsa,b,c Wing-Yan Manb, Bram Haversb Gert-Jan Keizerc

VU University
a b
Center for Advanced Studies Dutch Courage Foundation
c

Department of Computer Science IBM Netherlands Amsterdam, The Netherlands

Amsterdam, The Netherlands Amsterdam, The Netherlands treasurer@dutchcourage.org
hjsips@vu.nl {wingyanman,
bram_havers}@nl.ibm.com

Abstract— This paper presents an introduction to the Poseidon
Project. An open collaborative research project aiming to reduce
the Water Footprint in (crop) agriculture, by opening up access
to low-cost in situ sensing and cognitive computing to farmers
globally. Currently the Poseidon Project, is in the inception
phase. We give an overview of the motivation, the underlying
scientific principles, and a future outlook on the project.
Index Terms—Water Footprint, Open Computing, Cognitive
Computing, Machine Learning
I. INTRODUCTION
Climate Change is widely agreed to be already a reality [1].
The extreme weather in recent years [2][3], provides visible
proof of the changes in our global climate and the need to
mitigate the consequences. Countries in the developed world,
such as the Netherlands and the UK are preparing themselves
for increased risk of flooding from ocean and rivers [4], by
rapidly increasing the available budgets for Flood Control.
In the developing world however, most severe impact is
expected in food security and access to clean drinking water
[1]. In many countries in Africa, Asia and Latin America,
where agriculture is the major source of income –e.g. 60-70%
of the population in Africa relies on agriculture for their
income- the impact of Climate Change is expected to be
potentially disastrous [5], strengthened by the fact that the low
access to technology in the developing world makes the
agricultural systems less resistant to the effects of Climate
Change [1].
Even more stress is expected on global agriculture, through
rising population and increasing welfare in many parts of the
world, changing both the amount of food necessary, as well as
the types of food demanded. In a recent review paper, Lal [5]
concludes:
(1) To accommodate the increasing world population,
combined with the growing preference for animal
based diets, global food production needs to increase
substantially, and
(2) Global soil and water resources are limited, vulnerable
and coupled with the very processes that govern
climate change.
Currently, many regions in the world are already seeing the
effects of the continuously increasing demand. A prominent
example is Central Asia, as observed by Sorg et al. [6]
“Water shortages in summer will place the entire region’s
agricultural system under pressure, thus fuelling tensions
that have existed since the collapse of the Soviet Union in
the early 1990s. The high water demand for irrigation has
already transformed downstream sections of powerful
rivers such as the Syr Darya, Amu Darya and Ili into small
rivulets, thus exacerbating the drying-out of the Aral Sea
and Lake Balkash.”
Research and policy have traditionally observed shortages
in water as national or regional problems. Hoekstra and
Mekonnen [7] observe that governments have traditionally
matched national water demand to national water supply,
studied at local, national or river-basin level, thereby ignoring
the global dimension of water demand patterns.
However, given the active global trade in water-intensive
commodities, such as cotton or bovine meat, they propose to
measure the Water Footprint of a nation, rather than its
national water use. Figure 1 depicts the Virtual Water balance
of nations, showing in red the countries which are high
importers of fresh water and in green the countries which
export fresh water.
By using this definition, Chapagain et al. [8] conclude that
EU25 consumers are 20% responsible for the decline of the
Aral Sea in Central Asia, the most famous example of the
effects of water extraction for irrigation.
All in all, mitigation of the global water shortage should
aim at 2 core elements:
Figure 1. Virtual water balance per country and direction of gross virtual water flows
related to trade in agricultural and industrial products over the period 1996–2005. Only
the biggest gross flows (>15 Gm3∕y) are shown. [6]
(1) Increasing the resistance of third-world agriculture
towards the effects of climate change and
(2) Increasing awareness about the global dimension of
water shortage.
In this paper, we describe the Poseidon Project, aimed at
using recent technological advances to tackle the first of these.
II. INCREASING RESISTANCE: PRECISION IRRIGATION
Precision irrigation is defined as “site-specific irrigation
management that relies on variable application of water”. It is
emerging as a potential solution to increase the productivity
and reduce the environmental impact of irrigated agriculture
[5]. Studies have demonstrated that using precision irrigation,
water-use by crop-agriculture can be reduced by up to 20% in
Australia [12], helping tremendously in lowering agricultural
footprint.
The situation is even more striking in the arid regions of
Central Asia, where excessive, non-beneficial water use as
deep percolation and an inadequate use of the available soil
water and groundwater contribution is commonly practiced.
Simulations [14] have demonstrated that water usage could be
easily reduced by 25%, through implementation of mild deficit
irrigation. It may be expected that more advanced models and
realtime, in situ monitoring could bring even more reduction in
water usage [15].
However, despite the promise of lowering water usage and
the obvious need for water-usage reduction in many parts of
the world, precision irrigation has not found global uptake as of
today. For instance, Roth et al. observe that in the Australian
Cotton Industry, one of the most advanced agricultural
industries in terms of irrigation-scheduling tools only ~40% of
production is aided by Soil-Moisture probes [13].
Vellidis et al. [15] observe 3 core reasons for a low uptake of
precision agriculture, current systems are either (1) too expensive, (2)
too unreliable or (3) too complicated for uptake in a farm
environment.
We believe that recent developments in Computer Science can
bring solutions to all of these blockades, enabling a leap-frog in
adaption of precision agriculture. In the remainder of this paper, we
will discuss how Cognitive Computing and Open Source hardware are
expected to contribute.
III. OPEN COMPUTING
Significant progress has been made in ubiquitous
computing. The rise of Open Source hardware, such as the
Raspberry Pi and Arduino, dramatically reduced the cost of
hardware deployment in, most prominently research and
education.
For instance, Fisher and Gould [16] report the use of Open
Source hardware in the context of Precision Agriculture
Research. In a comparison with commercial-grade hardware,
they conclude that Open Source hardware can be a reasonable-
quality, low-cost alternative for commercial solutions, which
often create a large cost boundary for in field research projects;
the reported Open Source soil moisture probe promises great
cost-reduction for Precision Irrigation; the reported hardware
solution costs around $100,-, where commercial, closed,
devices cost at least the tenfold.
However, they conclude that significant effort has to be
spent in the calibration of sensors and that the use of these open
solutions requires significant technical knowledge from the
users; moreover, the quality of low-cost sensors may lag
behind compared to commercial devices.
IV. COGNITIVE COMPUTING
Cognitive computing promises to have a disruptive effect
on computer science [9]. At its core are problems that humans
find simple to solve, and machines find difficult. Applications
include question answering, medical diagnosis, natural
language processing, etc. Recent projects, such as IBM’s
highly publicized Watson question answering computer, have
highlighted that new approaches and massive computing can
combine to make significant progress on these problems.
Example domains include like Astronomy [9], Medicine [10]
and Flood Control [4].
These machine learning based approaches bring a radical
change: rather than spending significant effort on programming
a solution, effort is spent devising evaluation metrics and
gathering annotated data – ground truth data – for training and
evaluating a system.
Similar approaches have been shown successful on (low-
cost) sensors, as demonstrated in the work done by Mititelu et
al.[19] and Overeem et al.[18] These former of these studies
demonstrates that, using Machine Learning, we may detect
normal and abnormal behavior in a levee, without making
assumptions of the underlying geophysical structure, thereby
reducing the reliance on mathematical modelling. The latter is
an example of Social Sensing, systems in which users are
encouraged to share data and/or Social Media sources are used
to learn about the physical world. In the work by Overeem et
al., the embedded sensors in Smart Phones can be correlated
with the ambient temperature and, as such, enable highly
granular climate measurements in an urban environment; a
highly relevant topic in Urban Water Management, where
current systems are to coarse in time and space to effectively
model the microclimate within a city.
Social Sensing is gaining uptake and application, yet
various authors have noted that correct analysis of Big and
human-generated data is often very difficult and error-prone.
Grimberg et al. [20] demonstrate this, by analyzing Twitter
and Foursquare data generated around Hurricane Sandy (2012).
They observer that interesting patterns can be derived, like the
the stop and start of nightlife before and after the storm and the
large increase in grocery shopping before Sandy. However,
using these same data, they also demonstrate that one could
easily draw wrong conclusions, for example that Manhattan
(the area with most active Twitter users) was the pinnacle of
the storm.
Boyd et al. [21] issue an even wider claim against the idea
that Big Data would solve issues of bias and descriptiveness,
often assumed in applications of Big Data. They suggest that
the analysis of Human Generated Big Data would require
interaction between Social Sciences and Computer Science, to
better deal with the underlying social phenomena found in
human generated data.
Recent work by Aroyo and Welty, CrowdTruth [17], aims
to tackle these issues, by harnessing the span of opinion (or
disagreement on the semantics of text and images, by means of
Crowd-Sourcing. It is based on two main observations; (1)
often, the interpretation of data is a matter of semantic
interpretation, multiple opinions may hold and may correct,
e.g. on the meaning of text and terminology and (2) humans
and computers can work together successfully, on tasks which
are difficult for computers, but easy for humans and vice-versa.
An example of the former is the understanding of language,
which is difficult for computers and easy for humans. The
latter can be exemplified by, e.g., large calculations; easy for
machines and difficult for humans. Currently,the Crowd-
Watson system applies CrowdTruth to generate training data
for IBMs Watson System and is focused on annotation of Text,
Images and Video [22].
V. THE POSEIDON PROJECT
The Poseidon Project aims to combine Open Source
hardware, CrowdTruth and Machine Learning to increase the
reliability of low-cost sensor measurements in the context of
precision agriculture At the heart of Poseidon is an Open
Source cognitive computing platform, based on IBM Bluemix 1.
Figure 2 depicts the core components of the platform,
which we will discuss in more detail below.
Within the in situ sensing layer. are the physical Open
Source measurement devices deployed in an agricultural field.
These devices consist of Open Source computing platforms
(like arduino and raspberry pi) and sensors, which may span a
wide variety of configurations and quality. The Open Devices
will connect to the data collection layer, using the open
MQTT 2 protocol.
1
http://www.bluemix.net
2
http://www.mqtt.net
Within the data collection layer, data from the remote
devices are collected and abstracted (to compensate for
differences in measurement unit, granularity, etc.). Moreover,
the data of sensors on close geographical and temporal distance
are collected in Measurement vectors (analogous to the
CrowdTruth Annotation Vectors), for details on CrowdTruth,
we refer to [17] and [22]. In short, using this paradigm, we
consider the measurements of groups of low-cost sensors as
opinions on the actual value, rather than individual
measurements.
In contrast to, e.g. the work by Mititelu et al. [19] where
unsupervised machine learning (Kohonen SOM) is used to
detect outliers, a common practice in anomaly detection; this
approach would allow for a evaluation of individual sensors
through the CrowdTruth metrics, as reported in [22], [23] and
[24]. The crowdsourcing concepts of worker and micro-task
are transferred in our approach to sensor and temporal unit, e.g.
a soil-moisture sensor – worker - completes one measurement
–microtask- per hour. Similarly, the CrowdTruth-metrics,
identified to detect, e.g. spam, can be transferred to sensory
data. Worker Quality becomes Sensor Quality (e.g. identifying
failing sensors). For a full overview of metrics, we refer to
[22]. Within the metrics layer, these CrowdTruth metrics are
applied to the vector representations of our sensory data.
On top of these metrics (which result in a normalized
numerical value) and the reported measurement vectors,
classifiers are trained, to identify 3 classes: (1) need for
irrigation, (2) no need, (3) over-irrigated. In the future the input
dimensions for these classifiers are expected to be extended
with, e.g. weather reports (one may choose not to irrigate, or to
irrigate extra) based on weather data.
Finally, within the presentation layer, results are reported
to the farmers by means of a mobile push message (through
app, or sms).
Figure 2. High level architecture for the Poseidon Platform.
VI. DISCUSSION
Within this paper we have presented a framework for
precision irrigation based on low-cost sensors. This framework,
exploits advances in Cognitive Computing and Open Source
hardware, to tackle the current inhibitors for large-scale
adoption of precision irrigation.
In the coming decades, rising population and consumption
are expected to place additional stress on the available natural
resources of the world. The main premise behind the Poseidon
Project is that we can no longer afford to perceive these
challenges within geographical boundaries (i.e. as a regional
problem), or within the boundaries of a scientific discipline
(i.e., hydrology problem). Solutions may be found by applying
advances of some fields on the issues of another field. One
example thereof is to apply CrowdTruth and the related metrics
in the context of low-cost sensors, yet, we may expect that
other approaches, such as the use of Crowd-Sourcing and
Niche-Sourcing could prove to be successful in Precision
Agriculture. We may expect to see applications in which
farmers can share and vote on pictures of crops to help their
peers to identify health and quality issues (niche sourcing) or to
have large crowds identify anomalies in crops (crowd
sourcing). These approaches have proven to be successful in,
e.g., astrophysics [25] and could potentially help to preserve
harvests of the future.
ACKNOWLEDGMENTS
We thank the Dutch Courage foundation for enabling this
project, and the volunteers at IBM and Delft University of
Technology, who generously offer their spare time in the
development of this platform.
REFERENCES
[1] Kurukulasuriya, Pradeep; Rosenthal, Shane. (2013). Climate
Change and Agriculture : A Review of Impacts and Adaptations.
World Bank, Washington, DC.
[2] Lal, R.; Delgado, J.A.; Gulliford, J.; Nielsen, D.; Rice, C.W.;
van Pelt, R.S. Adapting agriculture to drought and extreme
events. J. Soil Water Conserv. 2012, 67, 162–166.
[3] Lyall, S. Heat, Flood or Icy Cold, Extreme Weather Rages
Worldwide; The New York Times: New York, NY, USA, 2013.
[4] Sips, R., van der Vlis, A., Nagel, R. And Havers, B. (2013). A
Case for Evidence Based Levee Management using Sensor
Technology. IBM Journal of Research and Development. 57(5)
pp. 5:1-5:12.
[5] Lal, R. (2014). Climate Strategic Soil Management. Challenges.
2014, 5, 43-74.
[6] Sorg, A., Bolch, T., Stoffel, M., Solomina, O., & Beniston, M.
(2012). Climate change impacts on glaciers and runoff in Tien
Shan (Central Asia). Nature Climate Change, 2(10), 725-731.
[7] Hoekstra, A., Mekonnen, M. (2012). The water footprint of
humanity. Proceedings of the National Academy of Sciences,
109(3). 3232-3237.
[8] Chapagain, A., Hoekstra, A., Savenije, H. and Gautam, R.
(2006). The water footprint of cotton consumption: an impact of
the worldwide consumption of cotton products on the water
resources of cotton producing countries. Ecological economics.
60(1). 186-203.
[9] Kelly III, J., Hamm, S. (2013). Smart Machines: IBMs Watson
and the Era of Cognitice Computing. Colombia University
Press.
[10] Engbersen, A., Brink, A., and Sips, R. (2013). SKA: The
Ultimate Big Data Challenge. World Summit on Big Data and
Organizational Design. Sorbonne, Paris, 2013.
[11] Hoogendoorn, M., Moons, L., Numans, M. and Sips, R., (2014).
Utilizing Data Mining for Predictive Modeling of Colorectal
Cancer using Electonic Medical Records. Brain Informatics and
Health. Lecture Notes in Computer Science. 8609. pp. 132-141.
[12] Hedley, C., Yule, I. and Bradbury, S. (2010). Analysis of
precision irrigation for variable soils at five pastoral and arable
production sites in New Zwaland. In 19th World Soil Congress.
pp. 1-6.
[13] Roth, G., Harris, G., Gillies, M., Montgomery, J. and
Wigginton, D. Water-use effciciency and productivity trends in
Australian irrigated cotton: a review. Crop & Pasture Science,
2013, 64, 1033–1048.
[14] Pereira, L. S., Paredes, P., Сholpankulov, E. D., Inchenkova, O.
P., Teodoro, P. R., & Horst, M. G. (2009). Irrigation scheduling
strategies for cotton to cope with water scarcity in the Fergana
Valley, Central Asia. Agricultural water management, 96(5),
723-735.
[15] Vellidis, G., V.13 Garrick, S. Pocknee, C. Perry, C. Kvien, M.
Tucker. (2007). How wireless will change agriculture. In:
Stafford, J.V. (Ed.), Precision Agriculture ‘07 – Proceedings of
the Sixth European Conference on Precision Agriculture
(6ECPA), Skiathos, Greece, pp. 57-67.
[16] Fisher, D., Gould, P. (2012). Open-Source Hardware is a low-
cost alternative for scientific instrumentation and research.
Modern Instrumentation. 2012, 1, 8-20.
[17] Aroyo, L., Welty, C. (2014). Truth is a Lie: 7 Myths about
Human Annotation. AI Magazine. 2014 (in press).
[18] Overeem, A., R Robinson, J. C., Leijnse, H., Steeneveld, G. J., P
Horn, B. K., & Uijlenhoet, R. (2013). Crowdsourcing urban air
temperatures from smartphone battery
temperatures. Geophysical Research Letters, 40(15), 4081-4085.
[19] Mititelu, G., Sips, R., Havers, B., Szlavik, Z., (2012). Towards
an Universial Change Detection Framework in Leveers.
Proceedings of the 22nd BeNeLux Conference on Artficial
Intelligence., Maastricht, The Netherlands.
[20] Grinberg, N., Naaman, M., Shaw, B., & Lotan, G. (2013, July).
Extracting Diurnal Patterns of Real World Activity from Social
Media. In ICWSM
[21] Boyd, D., & Crawford, K. (2012). Critical questions for big
data: Provocations for a cultural, technological, and scholarly
phenomenon. Information, Communication & Society, 15(5),
662-679
[22] Inel, O., Khamkham, K., Cristea, T., Dumitrache, A., Rutjes, A.,
van der Ploeg, J., Romaszko, L., Aroyo., L., Sips, R. (2014).
CrowdTruth: Machine-Human Computation Framework for
Harnessing Disagreement in Gathering Annotated Data.
International Semantic Web Conference 2014. (in press).
[23] Aroyo, L., Welty, C.: Measuring crowd truth for medical
relation extraction. In:AAAI2013 Fall Symp. on Semantics for
Big Data. (2013)
[24] Soberón, G., Aroyo, L., Welty, C., Inel, O., Lin, H., Overmeen,
M.: Measuringcrowd truth: Disagreement metrics combined
with worker behavior filters. In: Proc. of CrowdSem2013
Workshop), ISWC2013. (2013)
[25] Lintott, C. J., Schawinski, K., Slosar, A., Land, K., Bamford, S.,
Thomas, D and Vandenberg, J. (2008). Galaxy Zoo:
morphologies derived from visual inspection of galaxies from
the Sloan Digital Sky Survey. Monthly Notices of the Royal
Astronomical Society, 389(3), 1179-1189.