You are on page 1of 12

Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture


journal homepage: www.elsevier.com/locate/compag

Original papers

AgroDSS: A decision support system for agriculture and farming



Rok Rupnik, Matjaž Kukar, Petar Vračar, Domen Košir, Darko Pevec, Zoran Bosnić
University of Ljubljana, Faculty of Computer and Information Science, Večna pot 113, Ljubljana, Slovenia

A R T I C LE I N FO A B S T R A C T

Keywords: Decision support systems, data analysis and data mining have become significant tools for improving business in
Decision support system professional world. The emerging technologies are making the precision agriculture omnipresent and allow
Agriculture potential for enriching it with computer-assisted decision support systems for farm management. In this paper
Farming we describe a novel system AgroDSS that bridges the gap between agricultural systems and state-of-the-art
2000 MSC: decision support methodology. The described system is intended for integration into the existing farm man-
68U35 agement information systems and provides a cloud-based decision support toolbox, allowing farmers to upload
68T05 their own data, utilize several data analysis methods and retrieve their outputs. The implemented tools include
predictive modeling with explanation, accuracy evaluation, time series clustering and decomposition, and
structural change detection. They can help users make predictions for simulated scenarios and better understand
the dependencies (interactions) within their domain. We apply the AgroDSS system on a case study of pest
population dynamics, illustrating the potential for its use.

1. Introduction designed as a decision support toolbox, where farmers can upload their
own data of interest, utilize several offered artificial intelligence (data
In mid 1990s, Heuvel (1996) wrote about precision agriculture – the mining and statistical) tools and retrieve their outputs. The output re-
big information upgrade in agriculture that was yet to arise. In his work sults are intended to provide users with automated decision support,
he emphasized that the emerging technologies are driving the devel- predictions for simulated scenarios, better understanding of the sub-
opment of agriculture, which will be able to evolve to a more efficient mitted data and explanation of the dependencies (interactions) within
level in terms of agricultural production and land stewardship. It seems the data.
that today we are in the era of making precision agriculture omnipre- As motivated by related work (Fountas et al., 2015a; Kaloxylos
sent, whilst trying to enrich it with computer-assisted Decision Support et al., 2014; Sørensen et al., 2011), AgroDSS is implemented as a cloud-
Systems (DSS) for entire farm management (with the goal of optimizing based service, which enables various software products for agriculture
returns on inputs while preserving resources). to facilitate decision support to users. As many other existing decision
The majority of marketed state-of-the-art software products for support systems, we aim AgroDSS to be a tool that can improve business
agriculture utilize various sensors to collect data, in contrast to history value through savings, avoiding risks, better planning and taking better
that involved manual data input (Kaloxylos et al., 2012; Sørensen et al., advantage of resources. As AgroDSS can be used with different problem
2010). Although this results in large amounts of stored data, data is domains, business benefits depend on each individual business case; we
most frequently analyzed only with basic statistical or visualization conclude the paper by summarizing such a case study.
methods. The latter neglects the potential for application of advanced The paper is structured as follows. In Section 2 we provide a his-
data analysis tools that stem from the fields of data mining. The reason torical overview of how the computers and decision support systems
for this mostly lies in the fact that the development and deployment of have been used in agriculture and farming. In Section 3 we describe the
advanced data mining methods demand cutting-edge knowledge and design, functionalities and the implementation of the novel AgroDSS
skills in understanding these methods. system. We proceed in Section 4 by providing a brief case study of pest
In this paper we describe a novel system AgroDSS that bridges the analysis in orchards and fields, which serves as a demonstrative proof of
gap between agricultural systems and state-of-the-art decision support concept. We conclude the paper in Section 5.
methodology. The described system is intended to provide computer
decision support to farmers in an approachable manner. The system is


Corresponding author.
E-mail addresses: rok.rupnik@fri.uni-lj.si (R. Rupnik), matjaz.kukar@fri.uni-lj.si (M. Kukar), petar.vracar@fri.uni-lj.si (P. Vračar), domen.kosir@fri.uni-lj.si (D. Košir),
darko.pevec@fri.uni-lj.si (D. Pevec), zoran.bosnic@fri.uni-lj.si (Z. Bosnić).

https://doi.org/10.1016/j.compag.2018.04.001
Received 14 November 2017; Received in revised form 20 February 2018; Accepted 1 April 2018
0168-1699/ © 2018 Elsevier B.V. All rights reserved.

Please cite this article as: Rupnik, R., Computers and Electronics in Agriculture (2018), https://doi.org/10.1016/j.compag.2018.04.001
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

2. Related work support system MAST to systematically summarize the data related to
mastitis; Kerr et al. (1999) developed a knowledge-based decision
Before the advent of precision agriculture, only a few works about support system for strategic planning on sub-tropical dairy farms. More
use of information technologies in farming or agriculture were pub- recently, Bryant et al. (2010) evaluated Farmax Dairy Pro that can
lished. Through the time, use of decision support systems, data analysis, predict annual yields (per cow and per hectare) for milk, fat, protein
and data mining have become more significant tools for improving and milk solids. Baudracco et al. (2012) proposed e-Cow that predicts
business in professional world. One of the earliest systems that provided herbage intake, milk yield and live weight change in dairy cows grazing
computer-assisted decision support was CropSyst (Stöckle et al., 2003). temperate pastures with and without supplementary feeding.
Its development started in early 1990s and evolved into a suite of Other examples of DSS for farming. Perini and Susi (2004) fo-
programs that include a cropping simulator, a weather generator, GIS cused on design issues faced during the development of a DSS to be used
(Geographic Information System) cooperator program, a watershed by technicians of the advisory service performing pest management
model, and several miscellaneous utility programs. According to the according to an integrated production approach. Abdullah and Hussain
website1, the developed model simulates the soil water budget, soil- (2006) put decades worth of pest scouting recordings, agriculture and
plant nitrogen budget, crop canopy and root growth, dry matter pro- metrological data into a novel Pilot Agriculture Extension Data Ware-
duction, yield, residue production and decomposition, and erosion, house and present some interesting findings through data mining using
using several parameters, such as cultivar selection, crop rotation, ir- an indigenous technique based on the crossing minimization paradigm.
rigation, nitrogen fertilization, tillage operations, and residue man- Xu et al. (2008) proposed a systematic approach based on integrated
agement. An important advantage of the contemporary systems is information systems for agricultural ecosystem management that ex-
availability in a cloud (Ampatzidis et al., 2016; Kaloxylos et al., 2014), tracts and integrates data about terrain, land use and planting. Sus-
which can ensure greater availability of the systems to users and ap- tainable agriculture was discussed by Kurlavičius (2009), where they
plications. proposed a DSS that predicts what kind of crops should be grown in
Crop simulations and predictions. Many such systems for simu- particular areas, what/how many animals should be kept and what/
lating crops through a myriad of tweakable variables were developed how much of resources are required under the environmental and other
and evaluated with different crops and climate conditions. In more conditions. Antonopoulou et al. (2010) proposed a Web-based Decision
recent publications, Abedinpour et al. (2012) performed an evaluation Support System to help farmers select the alternative crops and provide
of AquaCrop model for maize crop in a semi-arid environment. This the information about the cultivation period.
model attempts to balance accuracy, simplicity and robustness by using Cloud-based decision support services. Kaloxylos et al. (2012)
a relatively small number of explicit and mostly intuitive parameters surveyed the key functionalities of Farm Management Information
and input variables. These simulators can be successfully combined Systems (FMIS) and possibilities for their future improvement. They
with other models; for example, García-Vila and Fereres (2012) com- emphasized that farmers need the global market of services in the
bined AquaCrop with an economic model for the optimization of irri- cloud, which can collect joint intelligence that can be used to extend
gation management at farm level. Irrigation management had become their FMIS. Furthermore, the authors predict that the solutions/tools for
very important so dedicated models have spurred. Bazzani (2005) de- intelligence might in the future be located solely in the cloud, which
veloped a Decision Support System for Irrigation for short and long requires the intelligence to be distributed between the cloud and farms
term analyses of water, labor and machinery, considering different to compensate for lack of a stable Internet link in fields and rural areas.
types of soils, irrigation systems, water-yield functions and seasonality In their further work, Kaloxylos et al. (2014) presented a conceptual
trough multicriterial mathematical programming techniques; Paredes implementation of a cloud-based FMIS for managing a greenhouse.
et al. (2014) examined partitioning evapotranspiration, yield prediction Farmers who evaluated the FMIS characterized the use of services in the
and economic returns of maize under various irrigation management cloud as valuable due to possibility of selecting and changing a service
strategies. Today, soil moisture is analyzed with inferential frameworks provider, availability of decision support services, and open and stan-
that combine data to infer moisture, precipitation, temperature and dardized interfaces of the cloud FMIS that can enable cooperation
canopy coverage variation (Ghosh et al., 2014). Giusti and Marsili- among stakeholders.
Libelli (2015) proposed a fuzzy decision support system, intended to Fountas et al. (2015a) have explored future perspectives of FMIS
improve the irrigation based on given crop and site characteristics. It and advocate that cloud computing will be one of the key drivers in
comprises an inference system that outputs the best irrigation action to their future development. Nikkilä et al. (2010) and Tayyebi et al.
keep the soil moisture in the suitable bounds, resulting in significant (2016) argue that FMIS for precision agriculture needs access to various
water savings. systems and propose a concept of multi-tier architecture where any
In theory, the proposed models could be used with arbitrary crops, application can access FMIS data through web services exposed by
though in practice, many specific simulators for specific crops have FMIS. Sørensen et al. (2011) explore the concept of external assisting
been developed. For example, Mohanty et al. (2012) parameterized the services, such as ones that can be implemented in a cloud, to facilitate
APSIM model (Keating et al., 2003) for soybean and wheat crop of standardized data exchange formats, data semantics, and transforma-
subtropical central India; Sepaskhah et al. (2013) developed a dynamic tion of data into useful knowledge for decision making. Tan (2016)
yield and growth model for saffron under different irrigation regimes; brought to attention that the further development of precision agri-
Dreassi et al. (2014) dealt with semicontinuous skewed spatial data of culture requires advanced tools to process data and transform them to
grape wine production; Sun and Ren (2014) assessed crop yield and practical decision actions. He developed a cloud-based decision support
optimize water scheduling for winter wheat and summer maize using system, which allows easier extensibility, integration of data from
soil and water assessment tool. In order to cope with the vast variability various sources (in different formats and semantics) and centralized
of nature and the uncertainties of agricultural production, the concept control of field devices.
of crop biometrics was defined as the scientific analysis of agricultural
observations confined to spaces of reduced dimensions and known
position with the purpose of building prediction models (Rovira-Más 2.1. Decision support systems methods and approaches
and Sáiz-Rubio, 2013).
Dairy farms. Allore et al. (1995) made a data-driven decision In the related work, several different decision support models and
approaches, which are based on data mining, have been used. In the
following we summarize their most common occurrences in the litera-
1
http://www.bsyse.wsu.edu/CS_Suite/CropSyst. ture:

2
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

• Probabilistic and optimization models: For sow herding, Pla et al. shown its high predictive value: Oliveira et al. (2012) compared it
(2004) developed a decision support system based on an embedded with the multiple regression on a fire occurrence in Mediterranean
Markov model, while Pomar and Pomar (2005) developed a system Europe domain; Rodriguez-Galiano et al. (2012) evaluated its use
in Milord II for efficient early identification of sows having low for land cover classification; Tatsumi et al. (2015) used it for crop
prolificacy performance in commercial pig farms. Scott et al. (2015) classification of homogeneous landscapes.
proposed a method for selecting appropriate suppliers that considers • Bayesian models: Wang et al. (2012) developed a Bayesian hier-
stakeholder requirements. They integrated their basic method, archical model that naturally combines information from multiple
called Analytic Hierarchy Process–Quality Function Deployment monthly surveys measured on different temporal supports, including
(AHP–QFD), with constrained programming and validated it within field measurement surveys and two farmer interview surveys. Their
the bioenergy industry. Thomopoulos et al. (2015) proposed an results indicate that the hierarchical model produces superior fore-
argumentation-based decision support system for agri-food chains casts to both the panel of experts and the composite estimator of the
that aims to optimize the nutritional, sensorial, practical and hy- mean of a normal distribution, while providing an accurate measure
gienic value of the food quality. Their system incorporates backward of uncertainty.
logical reasoning, starting from the goal, utilizing knowledge elici- • Other classification models: The paper of Waheed et al. (2006)
tation from domain experts or scientific papers. Although the au- investigates the potential of hyperspectral remote sensing data for
thors conclude that the task of knowledge modeling cannot be au- providing better crop management information for use in precision
tomated, they emphasize the importance of argumentation and farming. In the study, the ability of the classification and regression
modeling in decision support systems. trees (CART) decision tree algorithm were examined to classify
• Supervised learning models: It is worth mentioning that the well- hyperspectral data of experimental corn plots into categories of
known data mining software WEKA (Witten et al., 2016) started as a water stress, presence of weeds and nitrogen application rates. Li
project, funded from 1993 by the New Zealand government, aimed et al. (2012) were modeling maize irrigation with WOFOST simu-
to create a data mining workbench for contributing also to the lator, they performed an additional analysis of uncertainty that re-
agricultural industries (Hall et al., 2009). WEKA today features an vealed new insight into the risks associated with individual deci-
abundance of supervised learning algorithms, used by a wide re- sions. Du et al. (2013) presented a new algorithm for forecasting
search community. For example, Ramesh and Ramar (2011) used demand for perishable farm products, based on the support vector
WEKA to be the first to evaluate data mining techniques for Tamil machine. To fit a Gamma model, to get small area estimates and to
Nadu (state of India) soil data sets. They compared the performance evaluate their precision, Dreassi et al. (2014) used a hierarchical
of the decision trees, naive Bayes classifier and random forests and Bayesian approach in their Small area estimation for semicontin-
achieved 100% classification accuracy with the latter two. uous skewed spatial data (an application to the grape wine pro-
Today, many other implementation of classification and regression duction in Tuscany).
algorithms have been applied in various application areas; we pro- • Time series analysis: Michel and Makowski (2013) compared the
vide some typical examples below. Cortez et al. (2009) used the performance of seven basic statistical models and the dynamic linear
support vector machines (SVM) to outperform neural networks and model, for analyzing yield time series and prediction of yield trends.
multiple regression for predicting the human wine preferences They visualized yield increase rates obtained with a dynamic linear
(quality of wine) from objective physicochemical analytical tests. model for wheat from 1960 to 2010.
Their model revealed, which variables can be controlled during the • Genetic programming: Alhadidi et al. (2012) proposed and eval-
production, providing a direct insight into ways for improving wine uated a mathematical model that describes the reported data on
quality. Rumpf et al. (2010) tackled the problem for an early de- crop pests to get an accurate prediction of production costs, food
tection of sugar beet diseases using the SVM. They achieved the safety, and the protection of the environment. Their proposed ap-
classification accuracy of up to 97% for differentiation of healthy vs. proach is based on genetic programming and is explicitly directed at
diseased leaves and classification accuracy between 65% and 90% solving the symbolic regression of crop pest forecasting. Forecasting
(depending on the type and stage of the disease) for early detection agricultural output with an improved grey forecasting model based
of diseases. Pradhan (2013) compared the prediction performance of on the genetic algorithm was performed by Ou (2012).
decision trees (DT), support vector machine and adaptive neuro-
fuzzy inference system (ANFIS) for landslide susceptibility problem Rose et al. (2016) surveyed the available decision support tools in
in Malaysia using aerial photographs. They built several models the UK and discovered that 49% of farmers used some kind of decision
from different input attribute sets and achieved good predictive support tool to inform decisions. They determined that effectiveness of
performance, especially with DT and ANFIS. decision support is influenced by fifteen following characteristics: per-
Irmak et al. (2006) applied the back-propagation neural network to formance, ease of use, peer recommendation (encouragement of
predict the spatial distribution of soybean yields and to understand knowledge exchange), trust, cost, matching to farmers’ existing habits,
the causes of yield variability. The goal of Panda et al. (2010) was to relevance, targeting of advisers, user age compatibility, scale of busi-
investigate the strength of key spectral vegetation indices for corn ness, farming type, IT education, infrastructure (internet, devices,
crop yield prediction using neural network techniques. They showed workflows), compliance with market and legislation, and level of
that the corn yield was best predicted (with accuracy of marketing.
83.50–96.04%) using models that used the means and standard Based on the surveyed work we conclude that decision support
deviations of perpendicular vegetation index grid images. Karimi systems have been extensively studied for various aspects of agriculture
et al. (2006) compared the performance of support vector machine and farming. The related work shows an increasing demand for (1)
with artificial neural networks (ANN) for classification of hyper- predictive modeling, (2) data understanding and (3) statistical data
spectral images of corn fields in order to predict nitrogen application analysis, which motivates us to propose a system that can provide a
and weed infestation rates. They achieved favorable results, simplified utilization of such algorithms. In regard to positive char-
showing lower misclassification rates of SVM for that task than those acteristics of random forests (Oliveira et al., 2012; Rodriguez-Galiano
of ANN. et al., 2012; Tatsumi et al., 2015), which include the model’s non-
In the last decade, the advantages of the random forest classifier parametric nature, high predictive ability, internal evaluation of attri-
have been emphasized, including its non-parametric nature, high butes, robustness to noise, and avoidance of overfitting, we decide to
predictive ability, internal evaluation of attributes, robustness to implement it as a predictive model within AgroDSS. We intend to ad-
noise, and lack of proneness to overfitting. Several researches have dress at least six factors of effectiveness, given by (Rose et al., 2016)

3
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

(performance, ease of use, matching to farmers’ existing habits, re- mining process: business understanding, data understanding, data
levance, farming type) that can be easily addressed technologically and preparation, modeling, evaluation and deployment. The CRISP-DM
by offering non-domain-specific data mining algorithms. We present the process model therefore breaks down the life cycle of a data mining
chosen data mining algorithms and the proposed system in the fol- project into the six phases, which all include a variety of tasks. The
lowing sections. relation of CRISP-DM phases to the AgroDSS workflow steps is depicted
in Fig. 2. The workflow steps (step 1 – dataset creation, optional step 2 –
3. Decision support system AgroDSS dataset update, step 3 – quick data analysis, step 4 – training a pre-
dictive model, step 5 – model evaluation, and step 6 – predictive ana-
In the recent years, farm management information systems (FMIS) lysis) typically comprise one or more CRISP-DM phases. CRISP-DM
have evolved from simple farm recordkeeping into sophisticated and phases are related to AgroDSS workflow steps (Fig. 2) as follows:
complex systems to support all aspects of production management
(Fountas et al., 2015b). The purpose of modern FMIS is to manage 1. Business (problem) understanding: focuses on understanding the
demands to reduce production costs, comply with agricultural stan- project objectives from business (user’s) perspective and trans-
dards, and maintain high product quality and safety by ensuring access forming it into a data mining problem (domain) definition. At the
to timely information, elaborate estimation, simulation and decision end of the phase the project plan is produced. In AgroDSS, this phase
making tools. Decision support system that integrates all of the above is is a precondition for step 1 (users – farmers – should know what
therefore a crucial step in both day-to-day and strategic decision their farming problem is).
making. 2. Data understanding: starts with an initial raw data collection and
In this paper we describe a novel decision support system for agri- proceeds with activities such as getting familiar with data, dis-
culture and farming, called AgroDSS. The decision support aspect of the covering first insights into the data and identifying possible data
proposed system is based on data mining approaches, which can extract quality problems. In AgroDSS, this phase is a precondition for step 1.
useful information from large volumes of data. Data mining uses ad- In order to utilize the AgroDSS Web user interface, the user (farmer)
vanced algorithms to analyze the data and identify trends, distributions should possess a reasonable amount of domain and technical
and find patterns that can be hard to see with the naked eye. The newly knowledge. He should be able to define the decision problem (e.g.,
acquired knowledge can help better understand the data, see how data how to improve cattle feeding in order to increase milk production),
changes in time and help predict the future impact of different actions. identify relevant collected data (e.g., feed composition, daily milk
The modular architecture of AgroDSS allows easy (generic) integration production), format and upload the data, and review the results.
of API services into existing FMIS, and also exposes a simplified Web However, in a real-life situation the AgroDSS services will be in-
user interface (UI) to the end-user (farmer). While the integration with tegrated into a FMIS that will handle the problem definition (if not
web services is intended as a primary use of AgroDSS, the Web UI only pre-defined) and all the data handling process.
represents an additional demonstration/evaluation tool. The main 3. Data preparation: covers all activities to construct the final data set
production use of the AgroDSS will therefore originate from within from the initial raw data including selection of data, data cleaning,
FMIS, where data is to be prepared and submitted for evaluation, and the construction of data, the integration of data and the data for-
results presented without unnecessary user interaction. The inter- matting. In AgroDSS, this phase corresponds to steps 1, 2 and 3
mediate purpose of the AgroDSS will be to model the data and evaluate (creation of a dataset and partial uploads of data, as well as basic
the produced models. data analysis – descriptive statistics). While the data preparation
Although AgroDSS utilizes data mining algorithms, it is designed in step is mandatory and time-consuming for the demonstrational
a way that is approachable also to non-experts in the field. While in- AgroDSS Web interface users, this will be different in practice – the
tegration of API services into existing FMIS requires a programmer’s data will be collected and preprocessed in FMIS, and after selecting
knowledge (only during the integration), the farmers are the primary the required analysis, the data preparation phase will commence
users that should provide the domain knowledge prior to the integration, automatically, followed by communication using the AgroDSS API
i.e., define what the problem for the analysis is (e.g., prediction of a services.
certain phenomenon, what is a target prediction variable, change de- 4. Modeling: covers the creation of various data mining models. The
tection, seasonality analysis etc.). After the integration, the use should phase starts with the selection of data mining methods, proceeds
become routine, allowing farmers to run and understand the analyses with the creation (training) of data mining models and finishes with
with no difficulty. the assessment of the models. Some data mining methods have
We approached planning, design and implementation of the specific requirements on the form of data and the step back to data
AgroDSS system by considering the well-known Cross-industry standard preparation phase is often necessary. In AgroDSS, this phase corre-
process for data mining (CRISP-DM) (Chapman et al., 2000; Shearer, sponds to step 4 (depending of the problem type, users can select
2000), illustrated in Fig. 1. This standard encompasses good practices appropriate data mining methods).
for applying data mining techniques and defines the phases of data 5. Evaluation: evaluates the data mining models created in the mod-
eling phase. The aim of the model evaluation is to confirm that the
models are of high enough quality to achieve the business objec-
Business Data
tives. In AgroDSS, this phase corresponds to step 5 (depending of the
understanding understanding problem type, users can visualize evaluation reports for their data-
sets).
6. Deployment: comprises of activities to organize knowledge gained
Data
Deployment DATA through data mining models and present it in a way users can use it
preparation
within their decision making workflow. In AgroDSS, this phase
corresponds to step 6. The selected trained model is stored for future
Modeling re-use, and can be retrieved and utilized at any time. The storing of
Evaluation
the model is performed by serializing an R model object (R Core
Team, 2014), which is returned after the model’s generation, and
saving it as a string into the database.
Fig. 1. Phases of the cross-industry standard process for data mining (CRISP-
DM) that was the guideline for the development of the AgroDSS system. We designed the AgroDSS system with the following main goals:

4
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

CRISP-DM:
CRISP-DM: prediFtive
1 - business understanding CRISP-DM:
2 - data understanding
3 - data preparation model 5 - evaluation
evaluation

2. APPEND data 5. EVALUATE a


to existing dataset predictive model

1. create a 4. TRAIN a
START: new dataset
datasetready
predictive model
prediFtive CRISP-DM:
de¿netKeproElem
FolleFttKedata IorproFessing modelready 4 - modeling

3. run data 6. compute


ANALYSIS PREDICTIONS

dataanalysis prediFtive CRISP-DM:


results analysisresults 6 - deployment

Fig. 2. Workflow within the AgroDSS System. The boldface text denotes the CRISP-DM steps and text in the rounded rectangles denotes the AgroDSS workflow steps.
The latter are: dataset creation (step 1), update (optional step 2), quick data analysis (step 3), training a predictive model (step 4), model evaluation (step 5) and
predictive analysis (step 6). Steps within the workflow relate to corresponding phases of the CRISP-DM model.

• We aimed to implement a flexible and easy-to-use system that can be represent attributes (different descriptives of the current example).
also used by non-experts from the data mining field. Since the input data may contain missing values and many data
• The data mining tools should work with data from various domains; mining algorithms require a full data table, the preprocessing op-
the system should therefore allow users to upload and process their tions allow automatic imputation of missing values. For imputation,
own data. Minimal requirements about the data format (tabular the k-nearest neighbors algorithm (with k = 10) was used, which
form) should be defined. utilizes the most similar examples to deduce and fill-in the missing
• The implemented collection of data mining tools should be general values.
enough to be used with various problem domains. It should cover 2. Predictive modeling with explanation.
analyses from diverse subfields of data mining, such as supervised (a) Predictive modeling function builds a data mining model from
learning (predictive modeling), unsupervised learning (clustering), a given dataset (training dataset) that was uploaded to the
and time series analysis. server. The resulting data model can be afterwards used to make
• Users should receive numeric or visual feedback with the results of predictions for examples in another dataset from the same
their queries. problem domain (testing dataset). To model the data, random
• The system should be online and easily accessible to users. forest prediction model was used. Along with modeling the
• The system’s tools should be accessible by other information systems training set, the model performance is also estimated using the
through an application programming interface (API). out-of-bag (OOB) error that is assessed internally as the random
forest is being constructed. Each tree in the random forest is
The final architecture of the AgroDSS system is illustrated in Fig. 2. namely trained using a different bootstrap sample (leaving out
In step 1, the data for the analysis needs to be uploaded (in CSV format, about one third of the examples) and its performance is un-
target/dependent variable in the first column) to be made available for biasedly estimated on the left-out examples. The system auto-
the data mining algorithms. To do this, the users create and name a new matically selects the appropriate performance measure, de-
dataset. At any later time, data can be optionally appended to existing pending on the type of the target variable – classification
uploaded datasets (step 2), which is useful if we are dealing with per- accuracy for classification problems (discrete target variable)
iodic data collections (e.g. daily measurements). Afterwards, users can and the mean squared error for regression problems (continuous
either run any of the quick data analyses (step 3; clustering, structural target variable). Additionally, the random forest model also
analysis and sequence decomposition) or train a predictive model (step outputs the importance of individual attributes (estimated by
4) that is required for making predictions. Users can either evaluate the the accuracy decrease given different attribute permutations
accuracy of the predictive model on training or on independent test and by the Gini index decrease after a node split over all trees)
data (step 5) or use model to make predictions (step 6) for new, yet for predicting the outcome, which can be useful for post hoc
unseen data or for the simulated scenarios. feature selection performed by the user in FMIS.
(b) Prediction computation function for computing predictions
3.1. Decision support functionalities for testing dataset or new examples.
(c) Explanation of the learned knowledge in the model that
Following the goals, described in the previous section, we im- computes contributions of examples’ attribute values and returns
plemented the following data mining function within the AgroDSS its visual representation. The computed contributions represent
system: the explanation of the learned knowledge within the prediction
model (Štrumbelj et al., 2009; Štrumbelj and Kononenko, 2010).
1. Data upload and preprocessing. The function provides the farmers The absolute size of a contribution value is an indicator of the
to upload their data, which subsequently becomes stored on the attribute’s influence on the final decision, while its sign signifies
server. The uploaded data is required to be in tabular form, where whether the particular attribute value influenced the class
the rows represent individual observations (examples) and columns probability in a positive way (by raising the predicted

5
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

probability of the class) or negative way (by lowering the pre- The available test datasets include: iris (prediction of the Iris flower
dicted probability of the class). As such, they explain the im- species, Anderson (1935)), beavers (body temperature series of two
portance of different attributes and may reveal new knowledge beavers, Reynolds (1994)) and chickwts (chicken weights by feed type,
about the problem domain. McNeil (1977)).
(d) Predictive accuracy analysis plot that shows the dependency The REST API includes an option to adjust the randomForest mod-
of the predictive error on the number of trees, returned from the eling to specific use cases by providing alternative settings for default
random forest prediction model. values of parameters: ntree (the number of trees to grow), mtry
3. Time series analysis: (number of randomly sampled variables at each split), and maxnodes
(a) Time series clustering that is intended for discovering simila- (maximum number of terminal nodes).
rities within a dataset and represent them in separate data The services, implemented as Enterprise JavaBeans (EJB), can be
groups. To implement this function we used the Permutation used through REST API3. Its main request/response interface is de-
Distribution Clustering method (Brandmaier, 2015) that is tai- scribed in Table 1, and the user interface of the developed system is
lored to clustering time series. The clustering method finds si- shown in Fig. 4.
milar time series by observing differences between data per-
mutation distributions. The method is robust and suitable for
4. Case study: Population dynamics of the pests in orchards and
time series with changing distributions (i.e. containing concept
fields
drift) and with changing lengths.
(b) Time series decomposition function decomposes a time series
After the development of the system, a small case study was per-
into seasonal, trend and irregular components using moving
formed on data that were obtained from one of the partners (EFOS
averages. The function is intended for detection of periodic as
d.o.o., Slovenia) that collaborated within the EU-funded project AgroIT.
well as extraordinary events in the time stream. The decom-
The goals of the AgroIT project were to implement a cloud platform for
position was peformed using function decompose within statis-
integration of existing applications for farming, and to facilitate the
tical package R (R Core Team, 2014).
pilot use of those applications. As a part of the AgroIT, four pilot pro-
(c) Structural change detection that is intended to detect abrupt
jects were initiated in four EU countries, including the TrapView4
changes in dynamics of time series. Such function can be ben-
platform that was implemented by EFOS. TrapView enables pest mon-
eficial in case of population or parameter modeling, where
itoring in vineyards and orchards through pest traps. Pests are lured
alarms need to be raised in case of big data changes. For im-
into pest traps, where their photos are taken with a certain frequency
plementation of this function we utilized the strucchange library
and transmitted to TrapView servers over mobile network for analysis
(Zeileis et al., 2003) in statistical package R (R Core Team,
and pest recognition.
2014).
Our case study was performed on pests data, collected by TrapView.
The objective of the case study was to briefly demonstrate AgroDSS
3.2. System architecture functionalities, illustrate the system’s use, and provide examples of the
obtained results – thus, to serve as a proof-of-concept. The data includes
We implemented the data mining infrastructure using the WildFly measurements on various locations in four European countries between
Application Server2, custom R server and PostgreSQL DB Server as a May 2014 and December 2015. Each measurement includes a descrip-
data persistence layer. The system architecture (shown in Fig. 3) allows tion of the location (country, spatial coordinates and location type that
use of the data mining services via a REST API. The core of data pro- can be one of: field, forest, orchard or vineyard), a description of the
cessing and computation is implemented in R – a statistical software measuring device (model and device id), the timestamp of the mea-
package suited for data mining tasks (R Core Team, 2014). For the surement, the type of measurement (automatic, manually reviewed),
purposes of development and service exploration we have also im- the count of trapped pests, and the type of pest species. A sample of the
plemented an HTML web interface. We used the following R libraries: collected data is shown in Fig. 5. One of farmers using TrapView was
tackling the problems of predicting the pest population size, its dy-
• RCurl: for connecting to external data sources, namics and its potential anomalies (e.g. trap failure, maintenance
• RPostgreSQL: for connecting to the PostgreSQL database, problems). To address these goals, we decided to perform time series
• jsonlite: support for JSON data format, clustering for stream similarity detection (functionality 3a from Sec-
• cluster: for data clustering, tion 3.1), time series decomposition (functionality 3b), and struc-
• strucchange: for finding structural changes in time series, tural change detection (functionality 3c). We found that the re-
• pdc: for clustering time series data, maining functionalities were not relevant for analysis in this problem
• randomForest: for classification and regression using random forests. domain; some of their use is, however, illustrated in the screenshot of
the system in Fig. 4.
The database provides storage for users’ data, requested tasks, and Changes in the population of pests in each single location represent
their results. Task execution is performed asynchronously, guided with a time series, which we uploaded to the AgroDSS system and analyzed.
task queuing and facilitated using the Wildfly application server (EJB A visualization example of such time series (data stream) is shown in
Message Driven Beans). When user requests a data analysis, only meta- Fig. 6. The data refer to the Plum fruit moth captured in an orchard in
data (name of the dataset, analysis type and other parameters) are re- Poland. Each point represents the maximum value of the measurements
trieved from the database and sent to the R data mining server. The R from that day.
server afterwards requests the actual dataset contents and runs the Performed time series clustering enabled us to perform stream si-
analysis. The results are saved into the database and are available to be milarity analysis, as shown in Fig. 7. The clustering dendrogram (tree)
presented to the user, who can retrieve them using a task’s unique ID. shows separation of all available data streams into several groups ac-
The database also stores data from autonomous weather stations (in cording to their mutual similarity. By observing similarities one can
real time), provided by one of the AgroIT project partners (see Section quickly deduce that they differ in the pest population dynamics. Two
4), and sample test benchmark datasets. The weather data can be used sequences at the bottom (denoted with black lines) seem smoother than
to augment other data, if spatially proximate weather data is available.
3
The API is fully documented at http://agroit-dm.fri.uni-lj.si/docs/.
2 4
http://wildfly.org/. http://www.trapview.com/.

6
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

PostgreSQL
database
server

HTTP API
Data mining
services

R data
mining server

Fig. 3. Architecture of the implemented AgroDSS system.

Table 1
Interface of the implemented REST API.
Task Request Parameters Response

get the list of existing datasets GET/datasets/ / JSON encoded array of datasets

get the content of an existing GET/datasets/name name – name of dataset JSON encoded response containing the datasets
dataset name, size and CSV data

save a dataset POST/datasets/ name – name of new dataset JSON encoded response with name, size and
create csv – content of new dataset status

append data PUT/datasets/ name – name of new dataset JSON encoded response with name, size and
append/ csv – content of new dataset status

run a data mining task POST/analysis/run/ name – name of new dataset JSON encoded response containing status and
task – type of analysis (one of: build_model, eval_model, clustering, results (key-values or images)
structanalysis, seqdecomposition)
task_params – optional list of task-dependent parameters

the above four sequences (denoted with grey lines). By further ana- unseasonably early occurrence of specific pests) the domain expert (e.g.
lyzing the results of clustering, it turned out that each group represents from the company EFOS) could be informed as well, so that he could
different pest species – the four black sequences represent species Os- raise the alarm for a wider area or nationwide.
trinia nubilalis and two black sequences represent the Codling moth Fig. 9 shows the decomposition of a measurement sequence into the
species. This example demonstrates how clustering tools can be used to seasonal, trend and remainder (random) components. The trend com-
automatically detect similarities between independent data streams. ponent shows the overall movement in the measurements and offers the
Afterwards, users can further utilize this acquired knowledge to de- possibility of further analyzes (e.g. searching for causal relationships
termine the factors that affect the detected data similarities, or to au- between different factors and the observed trend within the pest po-
tomatically detect the anomalies in the data. An example of the latter pulation). The seasonal component (in this particular case, the basic
would be a detected sudden difference between expectedly similar data cycle represents one week) provides insights into processes that drive
stream (measuring points), which could indicate a fault in measuring the pests population. Within our case study, the analysis allows detec-
devices or procedures. tion of the increasing global trend of caught pest population over time,
Structural change detection enabled us to detect abrupt changes in as well as detection of the seasonal periodic cycles that correlate with
population dynamics. Fig. 8 illustrates points in time (dashed vertical weekly cycles. The sum of the trend, seasonal and remainder compo-
lines) where the underlying methodology detected such changes, while nents equals the original data series. The bars on the right hand side of
the horizontal lines represent the average values of measurements on the plot represent the same data range, allowing a relative comparison
individual segments. Such sudden and unexpected changes in the of the magnitudes of each component. Relatively low magnitudes of the
structure of sequences can represent an argument for automatic raising seasonal component (denoted by a high vertical bar) suggest that the
of alarms and notifying experts about the event that could alternatively weekly variability in the pest population is negligible.
also indicate a malfunction of the measuring equipment. For the users The farmers and analysts from the AgroIT project team reported that
of the AgroDSS Web user interface the alarms basically mean just sig- the use of AgroDSS brought a significant added value to TrapView
nificant structural changes in the time series. However, for a particular users. Although TrapView already contained its own decision support
problem, and within an appropriately configured FMIS with automatic module that is based on monitoring, it was not able to make predictions
data preparation and evaluation, it will be possible to daily check for for the future. AgroDSS supplemented the basic decision support with
significant anomalies and raise alarms in order to notify either the predictive functionalities that enabled farmers to effectively plan
farmer or the problem expert. In our specific case study of observing spraying according to the predicted population size of particular pest
pests population dynamics, the farmer could be notified of anomalies, species. For example, given the prediction that a particular pest will be
indicating an early need to replace the overfull sticky plate or the at the stage of deposition of larvae in three days at a particular location
pheromone bait. On the other hand, for some anomalies (such as of TrapView, a farmer is able to timely purchase and plan the use of a

7
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

Fig. 4. User interface of the AgroDSS system. The user interaction displays an example of generated data explanation (top), model performance (middle) and test data
output (below).

spray that is dedicated exactly for that particular stage of larvae de- data from their ERP systems. The data show that the use of the AgroDSS
position. positively influenced two key factors that contribute to a better fi-
Additionally, the project analysts also calculated the financial out- nancial outcome: (1) the increased crop quality, resulting in higher
comes due to the use of AgroDSS functionalities. Within the analyses, gained price of crop, and (2) the decreased spraying costs, caused by a
the data from the first two years of the AgroDSS use (2015 and 2016) lower number of needed sprayings. Note that although the cost of in-
were compared to the average data of three years preceding the in- secticide sprays slightly differed between farms and between years (the
tegration with AgroDSS (2012–2014). Table 2 displays the results for difference was less than 1.5%), the number of yearly sprayings de-
three farms (denoted with A, B and C) that were able to provide reliable creased from 6 or 7 to 4 or 5.

8
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

Fig. 5. A sample of the data about orchard pest count that was used for the case study.
120
100
80
Number of Pests
60
40
20
0

201í08í1 201í0í14 201í10í14 201í11í1 201í12í1


Date
Fig. 6. An example of a time series expressing pest population in a single location.

Fig. 7. Time series clustering, which automatically detected groups of different pest species.

5. Conclusion AgroDSS that provides a cloud-based decision support toolbox, allowing


farmers to upload their own data, utilize several data analysis methods
The emerging technologies are making the precision agriculture and retrieve their outputs. The implemented tools include predictive
omnipresent and allow potential for enriching it with computer-assisted modeling with explanation, accuracy evaluation, time series clustering
decision support systems for farm management. We developed a system and decomposition, and structural change detection. Within the short

9
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

20
Number of Pests
15
10
5

2015í06í04 2015í06í2 2015í0í16 2015í08í11 2015í10í04


Date
Fig. 8. Detection of structural changes in population dynamics (indicated by vertical lines).

Fig. 9. Decomposition of the measured data into


components. The decompositions allows better un-
25

derstanding of underlying patterns in a time series:


data

the trend component represents a long-term


15

change in the data, the seasonal component re-


presents a repetitive pattern with a fixed period,
and the remainder represents variations in the
5

data that remain unexplained after the other com-


ponents have been removed.
seasonal

00
í04
0
trend

20
10

0
remainder

í5
í15

5 10 15 20
time

Table 2
Data for three farms (denoted with A, B and C) for three years before using the AgroDSS (2012–2014) and first two years when the AgroDSS was used (2015 and
2016). The data reveals positive changes in two key factors (in boldface): crop price (reflected by crop quality) and spraying costs.
Average for 2012–2014 2015 2016

A B C A B C A B C

Crop per hectar [ton] 65,45 48,51 56,76 67,09 49,43 58,19 69,72 52,21 57,98
Number of hectars 8 15 24 8 15 24 8 15 24
Price earned for one tone of crop [€] 145 150 150 160 155 160 165 160 170
Crop revenue [€] 75.922 109.148 204.336 85.875 114.925 223.450 92.030 125.304 236.558
Cost for insecticide spraying [€] 1.960 3.150 5.040 1.400 2.100 4.200 1.120 2.100 4.020

10
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

case study we outlined potential benefits of the implemented tools that supplementary feeding. Animal 6, 980–993.
can enable knowledge extraction from the data and optimization of Bazzani, G.M., 2005. An integrated decision support system for irrigation and water
policy design: DSIRR. Environ. Model. Softw. 20, 153–163.
business. While the AgroDSS Web user interface demonstrates the sys- Brandmaier, A.M., 2015. pdc: An R package for complexity-based clustering of time
tem’s functionalities, the main use is provided by the REST API that series. J. Stat. Softw. 67, 1–23.
allows straightforward integration with existing FMIS and enables the Bryant, J., Ogle, G., Marshall, P., Glassey, C., Lancaster, J., Garcia, S., Holmes, C., 2010.
Description and evaluation of the farmax dairy pro decision support model. New
full use of AgroDSS services even for non-technical users. Zealand J. Agric. Res. 53, 13–28.
The CRISP-DM process model, as described in full in the paper, is Chapman, P., Clinton, J., Kerber, R., Khabaza, T., Reinartz, T., Shearer, C., & Wirth, R.,
not a practical solution for the majority of users, as the first three steps 2000. CRISP-DM 1.0 step-by-step data mining guide,.
Cortez, P., Cerdeira, A., Almeida, F., Matos, T., Reis, J., 2009. Modeling wine preferences
(problem understanding, data understanding and data preparation) by data mining from physicochemical properties. Decis. Support Syst. 47, 547–553.
require domain expert and data expert knowledge. These steps are to Dreassi, E., Petrucci, A., Rocco, E., 2014. Small area estimation for semicontinuous
remain firmly within the FMIS – a farmer will need to specify the skewed spatial data: an application to the grape wine production in Tuscany.
Biometrical J. 56, 141–156.
problem and data sources, the data will be prepared and submitted to
Du, X.F., Leung, S.C., Zhang, J.L., Lai, K.K., 2013. Demand forecasting of perishable farm
the AgroDSS service. AgroDSS therefore provides a tool to address the products using support vector machine. Int. J. Syst. Sci. 44, 556–567.
steps in the second half of the CRISP-DM that can be automated: Fountas, S., Carli, G., Sørensen, C.G., Tsiropoulos, Z., Cavalaris, C., Vatsanidou, A., Liakos,
modeling, evaluation, and the majority of deployment (by visually B., Canavari, M., Wiebensohn, J., Tisserye, B., 2015a. Farm management information
systems: current situation and future perspectives. Comput. Electron. Agric. 115,
presenting the learned knowledge). Once deployed, a model will easily 40–50.
be utilized from the FMIS. Fountas, S., Carli, G., Sørensen, C., Tsiropoulos, Z., Cavalaris, C., Vatsanidou, A., Liakos,
Within our further work we shall pursue the implementation of B., Canavari, M., Wiebensohn, J., Tisserye, B., 2015b. Farm management information
systems: current situation and future perspectives. Comput. Electron. Agric. 115,
additional functionalities within the AgroDSS system. The further de- 40–50.
velopment will include (1) development of additional predictive models García-Vila, M., Fereres, E., 2012. Combining the simulation crop model AquaCrop with
(such as naive Bayes, decision trees and neural networks) and ensemble an economic model for the optimization of irrigation management at farm level. Eur.
J. Agron. 36, 21–31.
learning (bagging), (2) possibility of model evaluation on independent Ghosh, S., Bell, D.M., Clark, J.S., Gelfand, A.E., Flikkema, P.G., 2014. Process modeling
test sets (including computation of evaluation measures, such as: sen- for soil moisture using sensor network data. Statist. Methodol. 17, 99–112.
sitivity, specificity and F1-measure for classification; (relative) mean Giusti, E., Marsili-Libelli, S., 2015. A fuzzy decision support system for irrigation and
water conservation in agriculture. Environ. Model. Softw. 63, 73–86.
squared error and (relative) mean absolute error for regression), and (3) Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., Witten, I.H., 2009. The
inclusion of feature selection methods (ReliefF, information gain, Gini weka data mining software: an update. ACM SIGKDD explorations newsletter 11,
index). Although currently implemented random forests allows good 10–18.
Heuvel, R.M.V., 1996. The promise of precision agriculture. J. Soil Water Conserv. 51,
accuracy and robustness, we shall implement the additional predictive
38–40.
models to offer additional tools that can improve either interpretability Irmak, A., Jones, J., Batchelor, W., Irmak, S., Boote, K., Paz, J., 2006. Artificial neural
(naive Bayes, decision trees) or predictive accuracy. Namely, the de- network model as a data analysis tool in precision farming. Trans. ASABE 49,
veloping field of deep learning using neural networks allows automatic 2027–2037.
Kaloxylos, A., Eigenmann, R., Teye, F., Politopoulou, Z., Wolfert, S., Shrank, C., Pesonen,
construction of new features that can improve capabilities of predictive L., 2012. Farm management systems and the future internet era. Comput. Electron.
modeling systems. Evaluation on independent test sets, additional Agric. 89, 130–144.
performance measures and feature selection methods shall allow users Kaloxylos, A., Groumas, A., Sarris, V., Katsikas, L., Magdalinos, P., Antoniou, E.,
Eigenmann, R., 2014. A cloud-based farm management system: architecture and
to evaluate data either prior to learning or longitudinally. They shall implementation. Comput. Electron. Agric. 100, 168–179.
enable greater control of the built model’s quality and signal the pos- Karimi, Y., Prasher, S., Patel, R., Kim, S., 2006. Application of support vector machine
sible need to repeat the learning process. Possibilities to implement technology for weed and nitrogen stress detection in corn. Comput. Electron. Agric.
51, 99–109.
other data mining algorithms are also open – however, this demand Keating, B.A., Carberry, P.S., Hammer, G.L., Probert, M.E., Robertson, M.J., Holzworth,
shall be driven by specific business requirements of users, who re- D., Hochman, Z., 2003. An overview of APSIM, a model designed for farming systems
present the crucial part of the AgroDSS development process. simulation. Eur. J. Agron. 18, 267–288.
Kerr, D., Cowan, R., Chaseling, J., 1999. DAIRYPRO – a knowledge-based decision sup-
port system for strategic planning on sub-tropical dairy farms. i. system description.
Acknowledgments Agric. Syst. 59, 245–255.
Kurlavičius, A., 2009. Sustainable agricultural development: knowledge-based decision
support. Technol. Econ. Dev. Econ. 15, 294–309.
The AgroDSS system has been developed within the AgroIT project,
Li, Y., Kinzelbach, W., Zhou, J., Cheng, G., Li, X., 2012. Modelling irrigated maize with a
which has been co-financed through the European union programme combination of coupled-model simulation and uncertainty analysis, in the northwest
CIP-ICT PSP. of China. Hydrol. Earth Syst. Sci. 16, 1465.
The authors thank EFOS d.o.o., Slovenia, for providing their col- McNeil, D.R., 1977. Interactive data analysis: a practical primer. Wiley.
Michel, L., Makowski, D., 2013. Comparison of statistical models for analyzing wheat
lected data and collaborating within the case study. yield time series. PloS One 8, e78615.
Mohanty, M., Probert, M., Reddy, K.S., Dalal, R., Mishra, A., Rao, A.S., Menzies, N., 2012.
References Simulating soybean–wheat cropping system: APSIM model parameterization and
validation. Agric. Ecosyst. Environ. 152, 68–78.
Nikkilä, R., Seilonen, I., Koskinen, K., 2010. Software architecture for farm management
Abdullah, A., Hussain, A., 2006. Data mining a new pilot agriculture extension data information systems in precision agriculture. Comput. Electron. Agric. 70, 328–336.
warehouse. J. Res. Pract. Inform. Technol. 38, 229–250. Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., Pereira, J.M., 2012. Modeling
Abedinpour, M., Sarangi, A., Rajput, T., Singh, M., Pathak, H., Ahmad, T., 2012. spatial patterns of fire occurrence in mediterranean europe using multiple regression
Performance evaluation of AquaCrop model for maize crop in a semi-arid environ- and random forest. For. Ecol. Manage. 275, 117–129.
ment. Agric. Water Manage. 110, 55–66. Ou, S.-L., 2012. Forecasting agricultural output with an improved grey forecasting model
Alhadidi, B., ALAFEEF, A., AL-HIARI, H., 2012. Symbolic regression of crop pest fore- based on the genetic algorithm. Comput. Electron. Agric. 85, 33–39.
casting using genetic programming. Turk. J. Electr. Eng. Comput. Sci. 20, 1332–1342. Panda, S.S., Ames, D.P., Panigrahi, S., 2010. Application of vegetation indices for agri-
Allore, H., Jones, L., Merrill, W., Oltenacu, P., 1995. A decision support system for cultural crop yield prediction using neural network techniques. Remote Sens. 2,
evaluating mastitis information. J. Dairy Sci. 78, 1382–1398. 673–696.
Ampatzidis, Y., Tan, L., Haley, R., Whiting, M.D., 2016. Cloud-based harvest management Paredes, P., Rodrigues, G., Alves, I., Pereira, L., 2014. Partitioning evapotranspiration,
information system for hand-harvested specialty crops. Comput. Electron. Agric. 122, yield prediction and economic returns of maize under various irrigation management
161–167. strategies. Agric. Water Manage. 135, 27–39.
Anderson, E., 1935. The irises of the gaspe peninsula. Bull. Am. Iris Soc. 59, 2–5. Perini, A., Susi, A., 2004. Developing a decision support system for integrated production
Antonopoulou, E., Karetsos, S., Maliappis, M., Sideridis, A., 2010. Web and mobile in agriculture. Environ. Model. & Softw. 19, 821–829.
technologies in a prototype DSS for major field crops. Comput. Electron. Agric. 70, Pla, L., Pomar, C., Pomar, J., 2004. A sow herd decision support system based on an
292–301. embedded Markov model. Comput. Electron. Agric. 45, 51–69.
Baudracco, J., Lopez-Villalobos, N., Holmes, C., Comeron, E., Macdonald, K., Barry, T., Pomar, J., Pomar, C., 2005. A knowledge-based decision support system to improve sow
Friggens, N.C., 2012. e-Cow: an animal model that predicts herbage intake, milk yield farm productivity. Expert Syst. Appl. 29, 33–40.
and live weight change in dairy cows grazing temperate pastures, with and without Pradhan, B., 2013. A comparative study on the predictive ability of the decision tree,

11
R. Rupnik et al. Computers and Electronics in Agriculture xxx (xxxx) xxx–xxx

support vector machine and neuro-fuzzy models in landslide susceptibility mapping Agric. 76, 266–276.
using GIS. Comput. Geosci. 51, 350–365. Stöckle, C.O., Donatelli, M., Nelson, R., 2003. CropSyst, a cropping systems simulation
R Core Team, 2014. R: A Language and Environment for Statistical Computing. R model. Eur. J. Agron. 18, 289–307.
Foundation for Statistical Computing, Vienna, Austria. Sun, C., Ren, L., 2014. Assessing crop yield and crop water productivity and optimizing
Ramesh, V., Ramar, K., 2011. Classification of agricultural land soils: a data mining ap- irrigation scheduling of winter wheat and summer maize in the Haihe plain using
proach. Agric. J. 6, 82–86. SWAT model. Hydrol. Process. 28, 2478–2498.
Reynolds, P., 1994. Time-series analyses of beaver body temperatures. Case Studies in Tan, L., 2016. Cloud-based decision support and automation for precision agriculture in
Biometry. pp. 211–228. orchards. IFAC-PapersOnLine 49, 330–335.
Rodriguez-Galiano, V.F., Ghimire, B., Rogan, J., Chica-Olmo, M., Rigol-Sanchez, J.P., Tatsumi, K., Yamashiki, Y., Torres, M.A.C., Taipe, C.L.R., 2015. Crop classification of
2012. An assessment of the effectiveness of a random forest classifier for land-cover upland fields using random forest of time-series landsat 7 etm+ data. Comput.
classification. ISPRS J. Photogramm. Remote Sens. 67, 93–104. Electron. Agric. 115, 171–179.
Rose, D.C., Sutherland, W.J., Parker, C., Lobley, M., Winter, M., Morris, C., Dicks, L.V., Tayyebi, A., Meehan, T.D., Dischler, J., Radloff, G., Ferris, M., Gratton, C., 2016.
2016. Decision support tools for agriculture: towards effective design and delivery. Smartscape: a web-based decision support system for assessing the tradeoffs among
Agric. Syst. 149, 165–174. multiple ecosystem services under crop-change scenarios. Comput. Electron. Agric.
Rovira-Más, F., Sáiz-Rubio, V., 2013. Crop biometric maps: the key to prediction. Sensors 121, 108–121.
13, 12698–12743. Thomopoulos, R., Croitoru, M., Tamani, N., 2015. Decision support for agri-food chains: a
Rumpf, T., Mahlein, A.-K., Steiner, U., Oerke, E.-C., Dehne, H.-W., Plümer, L., 2010. Early reverse engineering argumentation-based approach. Ecol. Inform. 26, 182–191.
detection and classification of plant diseases with support vector machines based on Štrumbelj, E., Kononenko, I., 2010. An efficient explanation of individual classifications
hyperspectral reflectance. Comput. Electron. Agric. 74, 91–99. using game theory. J. Machine Learn. Res. 11, 1–18.
Scott, J., Ho, W., Dey, P.K., Talluri, S., 2015. A decision support system for supplier Štrumbelj, E., Kononenko, I., Robnik-Šikonja, M., 2009. Explaining instance classifica-
selection and order allocation in stochastic, multi-stakeholder and multi-criteria en- tions with interactions of subsets of feature values. Data Knowl. Eng. 68, 886–904.
vironments. Int. J. Prod. Econ. 166, 226–237. Waheed, T., Bonnell, R., Prasher, S.O., Paulet, E., 2006. Measuring performance in pre-
Sepaskhah, A., Amini-Nejad, M., Kamgar-Haghighi, A., 2013. Developing a dynamic yield cision agriculture: Cart?a decision tree approach. Agric. Water Manage. 84, 173–185.
and growth model for saffron under different irrigation regimes. Int. J. Plant Prod. 7, Wang, J.C., Holan, S.H., Nandram, B., Barboza, W., Toto, C., Anderson, E., 2012. A
473–504. Bayesian approach to estimating agricultural yield based on multiple repeated sur-
Shearer, C., 2000. The crisp-dm model: the new blueprint for data mining. J. Data veys. J. Agric. Biol. Environ. Statist. 17, 84–106.
Warehousing 5, 13–22. Witten, I.H., Frank, E., Hall, M.A., Pal, C.J., 2016. Data Mining: Practical Machine
Sørensen, C., Fountas, S., Nash, E., Pesonen, L., Bochtis, D., Pedersen, S.M., ... Blackmore, Learning Tools and Techniques. Morgan Kaufman.
S., 2010. Conceptual model of a future farm management information system. Xu, L., Liang, N., Gao, Q., 2008. An integrated approach for agricultural ecosystem
Comput. Electron. Agric. 72, 37–47. management. IEEE Trans. Syst. Man Cybern. C (Appl. Rev.) 38, 590–599.
Sørensen, C., Pesonen, L., Bochtis, D., Vougioukas, S., Suomi, P., 2011. Functional re- Zeileis, A., Kleiber, C., Krämer, W., Hornik, K., 2003. Testing and dating of structural
quirements for a future farm management information system. Comput. Electron. changes in practice. Comput. Statist. Data Anal. 44, 109–123.

12

You might also like