Model-based decision support systems in Agriculture

This page is the result of a project, which are now concluded, and the page is no longer maintained. We have recently updated the links. Thus, we refer you to the current home pages of the people involved concerning on-going activities. If in doubt you are very welcome to contact the two authors of this page (Links). Computer-based decision support systems (DSSs) have a well-established tradition within agriculture. The DSS range from simple accounting-based systems to systems based on detailed determinstic or stochastic models. Clearly, using different methods to model the same domain will produce different results. Therefore, if we want to compare the strengths and short-comings of different DSSs, it is very important that the underlying models of the DSSs are well-documented. This is not always the case.

The purpose of this homepage is to focus on the underlying assumptions and model aspects of decision support systems used in agriculture. From this point-of-view links to other pages are established.

The intention of this homepage is to provide an open platform for links and discussion between researchers within the area of model-based decision support, event though we have used activities within the Dina collaboration as a starting-point. Furthermore, we will maintain a list of links to prototypes of the models under development, as well as links to applications outside agriculture.

Contents

News
Information about coming events and other news within the priority area
Model based decision support as a priority area
A description of the priority research area with selected references. Essential parts of the description are:
People
Links to people in Dina involved with model-based decision support systems in agriculture.
Links
Links to ongoing projects within agriculture and miscellaneous links to other areas of interest.
Links to prototypes
About this page

Previous news section

Dina workshop: Sequential Monitoring: Controlling Error rates. Research Centre Foulum. 20/3 - 2002, 9:45-16:00.

AFITA 2002 3rd AFITA Conference, Beijing, China, 2002. Afita homepage.

Model based decision support as a priority area

Essentiel parts of the model description are:

Observations and Measurements

All decision support systems are based on observations on the real systems. The method and detail of observation, however, differ widely. E.g. measurements of a sample of the system, measurement a the part of the system, that are at risk, automatic registration, video observations etc. Naturally, the quality of the support from the systems depends on these observations.

Examples of research in this area is ongoing work in the project Production Monitoring in Pig Herds.  Thomas N. Madsen where the use of measurements of water and feed consumptions are used for monitoring growth of piglets. Another aspect is the use of image analysis for weight assessment in pigs (Brandl, N. and Jørgensen, E., 1996).

Recognition of uncertainty in the system

When a model of a real-life system is constructed, uncertainty is usually an important factor. In the modelling process it is necessary to recognize and relate to this uncertainty. If, for example, it is decided to ignore the uncertainty of the system, the model builder must be aware of the risk of misinterpretations of the model results. The uncertainty of a system can be divided into the following types:
  • Uncertainty due to selection of wrong model
  • Uncertainty due to imprecise estimation of model parameters
  • Uncertainty due to imprecise estimation of the prior state of the system
  • Uncertainty due to stochastic processes in the the system
Usually some or all of these types of uncertainty are ignored. This section will classify the DSSs according to these criteria. Furthermore, links to areas, where the problems of ignoring the uncertainty is discussed, will be maintained.

Belief Management

This part of the decision support system is the part that handles estimation of parameters, learning, calibration etc. These different words all describe the process of converting observed values to model parameters.
Kalmanfiltering techniques and dynamic linear models.
These methods have been used in several studies. Thysen & Enevoldsen (1992) and Thysen (1992) used the method for monitoring fertility and health in dairy herds. Toft and Madsen (1998) used the method for investigating the trend and diurnal pattern of feed intake using data from automatic individual feeding equipment for slaughter pigs.
Bayesian networks.
The use of Bayesian Networks for belief management in agriculture is one of the main research efforts within Dina. In Thysen (1992) the technique has been used for building a probabilistic expert system for detection and diagnosis of mastitis. Rasmussen (1995) used a Bayesian Network for detection of parental errors using bloodtyping in connection with sire-evalution. Kristensen & Rasmussen (1997) used the technique for a DSS for Growing Malting Barley without use of Pesticides. Hansen & Riis (1999) used Bayesian networks to predict the optimal level of Nitrogen fertilization.Currently a system for reproduction monitoring and planning is being developed in the project Sow monitoring system.  A short introduction (in Danish) to the area can be found in Jørgensen & Lauritzen (1998). Small examples are described in Potential Application Areas for Bayesian Networks Within Animal Production. Recently a new blog Bayesian Networks and Markov Processes in Agriculture has been started with the intention of presenting these and newer examples
Other techniques.
An examples of another statistical method for handling the observations from the farm is found in Dethlefsen, C. & Jørgensen, E. (1996), where longitudinal data analysis is for estimation of the relation between parity and litter size in sows. The method takes the censoring due to culling into account.

Decision Support

The decison support aspects concerns issues, such as search algorithms for optimum, criteria for optimality, presentation of near optimal decisions, sensitivity of solutions to assumptions etc.
Monte Carlo simulation models.
The use of Monte Carlo simulation for herd analysis is implemented e.g. within the Dina pig simulation model and within SIMHERD (Simulation model of cattle herds).
Bayesian Network and Markov Chain simulation models.
Bayesian networks for decision support is used both for static and dynamic problem areas. Static problems investigated are e.g. the BOBLO system, the malt barley system and the mastitis detection problem. Dynamic problems has been modelled as markov chain models e.g. Jalvingh (1993). The use of bayesian network can be seen as a refinement of this method. Examples covers a system for prediction of herd lactation yield and the sow monitoring system.
Dynamic Programming.
Dynamic Programming. Dynamic programming or to be more specific Markov decision programming techniques are well established at least within animal husbandry (The replacement problem).  Examples of these techniques are the application of Multi-level hierarchich markov processes. In addition to the replacement problem, the techniques has been used for planning of delivery of slaughter pigs and optimal mating strategies.
Influence diagrams.
The influence diagrams can be seen as a combination of dynamic programming and the Bayesian networks. The technique has been applied to decision support for mildew management in winter wheat (Jensen, 1995) and is currently being explored for used within work planning during harvest. Recently tecniques for  e.g. search for near-optimal strategies has been developed.

Model evaluation

Obviously, it is a crucial part of the modelling process to compare the behaviour of the model with the real-life system it is modelling, and to adjust the model according to the evaluation. It is not trivial, however, due to the inherited uncertainty of the system. Two apparantly equal prior states of the system may result in different observed results, e.g. because of influence by stochastic processes. If the model ignores the uncertainty, it will come to the same calculated result for the two equal prior states, so at least one of the predictions is wrong.

If the model includes the uncertainty, the result will be a probability distribution, which may asign a positive probability to both observed results. In this case, the model must be evaluated by measuring how unlikely the true observations are according to the predicted probability distributions.

From prototype to end user

It seems to be a general experience that, even though large efforts are spent on development of DSS prototypes, most prototypes never come to a practical application. There may be several reasons for this, including the following:
  • The funding of research projects is focused towards development of prototypes, rather than operational systems.
  • The main purpose of the project was to achieve the knowledge and experience from developing the prototype.
  • The prototype requires input data or skills which can not be expected from the typical user
  • The prototype turned out to perform poorly
  • ...
If it is an intention of a research project to develop an operational DSS based on a prototype, it may be very important to be in close contact with the intended users of the system, in order to develop the system to fit the needs and the skills of the users, and to ensure that the necessary data input can be collected by the users.

Pl@nteInfo (http://www.planteinfo.dk/) is a web-based information and decision support system for crop production. This system has proved to be an efficient framework for (mainly simple) forecasting models for crop diseases and pests. Compared to PC-based DSSs where implementation and updating of DSSs are known to be costly and slow, internet based systems have a very short development line from researcher to end user. In addition, if the web-based DSS is a server-side implementation, the system developer may benefit from the input data supplied by users of the system.

Publications

Brandl, N. and Jørgensen, E. (1996).
Determination of Live Weight in Pigs from Dimensions by Image Analysis. Computers and Electronics in Agriculture 15, pp. 57-72.
Dethlefsen, C. & Jørgensen, E. (1996).
A longitudinal model for litter size with variance components and random dropout. Dina Research Report 50, Research Centre Foulum, P.O. Box 23, DK-8830 Tjele, Denmark.
Greve, Jørgen (1995).
The Dynamic Ranking Approach and Its Application in Piglet Produktion : Ph.D Thesis Faculty of Technology and Science at Aalborg University,  Dina Research Report no. 35, Dina Foulum, Research Center Foulum, 144 pp.
Hansen, M.D. & Riis, C. (1999).
Anvendelse af bayesianske netværksmodeller til analyse af kvælstofdata fra landsforsøgene. Speciale, Institut for jordbrugsvidenskab, KVL.
Jalvingh, A. (1993).
Dynamic livestock modelling for on-farm decision support. Ph.D. thesis. Department of Farm Management, Wageningen.
Jensen, A.L. (1995).
A probabilistic model based decision support system for mildew management in winter wheat, Ph.D. thesis, Dina Research Report no. 39, Dina Foulum, Research Center Foulum, 194 pp, PostScript file
Jørgensen, E. and S.L.Lauritzen (1998).
Bedre Beslutninger med Bayesianske netværk. Naturens Verden, 7, p 280-287.
Kristensen, A. R. and Jørgensen, E. (1996).
Multi-level hierarchic Markov processes as a framework for herd management support. Dina Research Report 41, pp. 1-29.
Kristensen, K. and I.Rasmussen (1997).
A decision support system for mechanical weed control in malting barley. Proceedings of the First European Conference for Information Technology in Agriculture, 447-452.
Kure, H. (1997).
Marketing management support in slaughter pig production.  108 pp.
Rasmussen, Lene Kolind (1995).
Bayesian network for blood typing and parentage verification of cattle, Ph.D. Thesis, Dina Research Report no. 38, Dina Foulum, Research Center Foulum, 172 + 18 pp.
Thysen, Iver (1992)
Monitoring bulk tank somatic cell counts by a multi-process Kalman filter , Dina Research Report no. 01, Dina Foulum, Research Center Foulum, 18 pp
Thysen, Iver (editor) (1992).
Proceedings of a Dina/MPP Symposium, Statistical and Expert System Methods in Matitis Control, Dina Notat no. 03, Dina Foulum, Research Center Foulum, 80 pp. Paper presented at Statistical and Expert System Methods in Mastitis Control, Research Centre Foulum, May 21, 1992.
Thysen, Iver and Carsten Enevoldsen (1992).
Visual monitoring of reproduction in dairly herds, Dina Research Report no. 10, Dina Foulum, Research Center Foulum, 16 pp.
Toft, N. (1998). The dynamic aspect of the reproductive performance in the sow herd. Dina Notat No. 70.
Toft and Madsen (1998).
Modeling eating patterns of growing pigs using dynamic linear models. Dina Research Report No. 69.
Rasmussen, Lene K. and Thiesson, Bo (1996).
Quantitative Adjustment of BOBLO, Institute for Electronic Systems, Aalborg University, 25 pp.
References to unpublished work
A Decision Support System for Growing Malting Barley without use of Pesticides. in prep.  Contact: Kristian Kristensen, DIAS
Sow monitoring system. Contact  Erik Jørgensen
Prediction of Herd Lactation Yield. Contact  Søren L. Dittmer Tools for Explanation in Bayesian Networks with Application to an Agricultural Problem
Small examples:  Potential Application Areas for Bayesian Networks within Sow Production
Multi-level hierarchic Markov processes. Contact  Anders R. Kristensen &  Nils Toft KVL
Work planning during Harvest. (Contact: Claus Grøn Sørensen, DIAS Bygholm )
Decision Support for Operational Health Management in Slaughter Pig Production. Contact  Erik Jørgensen

People

at Dina KVL

at Dina DJF

at Dina Aalborg

Links

Previous events

  • Dina Workshop (In Danish) Stokastisk modellering i planteavl - Problemstillinger og løsningsmetoder. Golf Salonen, Viborg, 31. oktober 2000. Presentations.

Projects

Decision support links

Bayes links

Links to prototypes

Currently there are no links to prototypes. Suggestions are welcome.

About this page

This page was maintained by Allan Leck Jensen (alj@agrsci.dk) and Erik Jørgensen (Erik.Jorgensen@agrsci.dk). Please send us updates, comments and information to be included on the page.
Dina logo Author: Erik.Jorgensen@agrsci.dk. Updated: September 2001 - links updated 2006