Home

CarottAge Windows pour les données Ter

Contents

1. choisir le Fichier source Annee du debut Col de filtre choisir le fichier de sauvegarde Figure 1 3 Le menu donn es M CarottAge Fichiers Donnees Configuration Hmm Apprentissage Diagramme de Markov Visualisation Help Valider la configuration Donnees Configuration Modele Diagramme Visualisation 0 1 2 3 4 cl 5 tournesol 6 colza 7 GH p d t 8 E pois 9 prair p 10 gt prair a 11 prair t 12 jachere id urbain 14 autres 15 amp indeterminee 16 Figure 1 4 Le menu configuration apr s avoir import teruti1 cfg M CarottAge DER Fichiers Donnees Configuration Hmm Apprentissage Diagramme de Markov Visualisation Help Creer le HMM lire le HMM mod Donnees Configuration Modele Diagramme Visualisation gt IK Figure 1 5 Le menu mod le Visualisation d un HMM ergodique une colonne de 2 tats 1 5 4 Le sous menu mod le Ce sous menu cf Fig 1 5 permet de construire les fichiers de descrip tion des HMM Il fera appel dans votre dos aux programmes model lin gen editmodel Deux familles de topologies sont possibles par l option cr er un mod le lin aire et colonne d tats La topologie choisie dans le fichier do_markov bat correspond une seule colonne de 6 tats Par d faut les tats sont associ s des lois uniformes E
2. 13 14 15 M Benoit J P Deffontaines F Gras E Bienaim and R Cosserat Agriculture et qualit de l eau une approche interdisciplinaire de la pollution par les nitrates d un bassin d alimentation Cahiers Agricultures 6 97 105 1997 M Benoit J L Fiorelli P Morlon and Y Pons Technical management a central point for agronomy challenge First European Congress of Agronomy Session V 01 Paris 1990 M Beno t and M C Muhar Farmers landuse and groundwater quality an interdisciplinary approach Congress Future of the Land Wageningen 1993 M Benoit and F Papy La place de l agronomie dans la probl matique environnementale Les dossiers de l environnement de l INRA 17 53 62 1998 D J Berndt Finding Patterns in Time Series In U M Fayyad G Piatetsky Shapiro P Smyth and R Uthurusamy editors Advances in Knowledge Discovery and Data Mining pages 229 248 AAAI Press The MIT Press 1996 J Boiffin E Mal zieux and D Picard Cropping systems for the future 3rd International crop science congress 17 22 august 2000 Hambourg Germany 2000 13 pages R Brunet La carte mod le et les chor mes Mappemonde 86 4 3 6 1986 F Burel and J Baudry Hedgerow network patterns and processes in france In Zonneveld and Forman editors Changing Landscape an ecological perspective pages 99 120 Springer 1990 P G Cox Some issues in the design of agricul
3. their dynamics and their spatial organizations focusing on the crop temporal rotations that are able to explain the risk of nitrate loss 52 The data mining software CAROTTAGE has been used on Ter Uti data from the Seine watershed Results are presented and analyzed below for a small district Crop sequences in Saint Quentinois The diagram shown in figure 2 7 displays the main annual transitions between crops and their evolutions The importance of the transition between two crops is expressed with the thickness of the line joining the two crops One can see that in this district the wheat based rotations are in a majority 27 12km 2km 6km 4km NS Bretagne a the basic grid France territory 250m 1500m c an air photography and its 6x6 grid Figure 2 6 Collecting the Ter Uti data 3820 meshes square France 4 air photographs are sampled in a mesh a 6x6 grid determines 36 sites 28 Set aside Potatoe Maize Rape seed 1992 1993 1994 1995 1996 1997 1998 1999 Minimal probability for displaying 1 5 Figure 2 7 Crop transitions between 1992 and 1999 in the district of Saint Quentinois North east of France Only the transitions whose probability is greater than 1 5 are displayed The question mark denotes the container state e the main transitions are wheat beet
4. 6 Visualisation des transitions entre cultures 11 1 5 7 Le sous menu visualisation Ce sous menu permet la visualisation des pdf associ es aux tats cf Fig 1 7 M CarottAge Fichiers Donnees Configuration Hmm Apprentissage Diagramme de Markov Visualisation Help nbrmax inPdf display Pdf a Donnees Configuration Modele Diagramme Visualisation Weight 1 0 L 1 ble Weight 1 0 1 1 orge Weight 1 0 1 1 colza Weight 1 0 1 1 mais Weight 0 9999999963 0 0 9078624845 0 907862 prairies_pp 1 0 9882453308 0 0803828 prairies_temp 2 0 9999999963 0 0117547 prairies_art Weight 1000000023 0 0 6648231745 0 664823 bois 1 0 7513660863 0 0865429 Sols_art_NB ra 0 8172837347 0 0659176 aut_sols_NANB 3 0 8500345238 0 0327508 jacheres 4 0 8785057161 0 0284712 pre vergers 5 0 9068580307 0 0283523 Sols_batis 6 0 9347942732 0 0279362 eaux 7 0 9533391837 0 0185449 zones_interdites 8 0 9716463387 0 0183072 potagers 9 0 9761042497 0 00445791 pruniers 10 0 9803838446 0 00427959 6_especes Figure 1 7 Visualisation des pdf du mod le ergodique La figure 1 7 m rite quelques explications Elle repr sente le HMM apr s apprentissage On remarque que les pdf associ es aux tats de Dirac sont rest es dans leur d finition initiale En revanche l tat container ini tialement loi uniforme s est peupl des occupations qui ne pouvaient tre capt es par les tats de Dirac Il s agit
5. Si on avait voulu travailler avec des triplets d occupation on aurait utilis teruti3 cfg L archive contient un fichier de commandes Windows do_example bat qui encha ne les commandes cr ation de la description du HMM le fichier 1in3 1st est cr e par la commande model lin gen exe 3 comme d crit page 22 Ce fichier d crit la topologie du HMM lin aire trois tats ainsi que les densit s de probabilit s pdf comme probability density function qui sont ici uniformes inventaire des observations le programme ter2indice tempo parcourt le fichier short example txt afin d inventorier toutes les observations possibles L inventaire est repr sent par la liste bin terutil 1st cr ation du Hmm le programme editmodel cr e le HMM partir des fichiers 1in3 1st et bin teruti1 1st estimation du Hmm le programme fwtInra joue le r le de la com mande estimate voqu e page 22 visualisation le programme gviewmod construit le fichier lin3 txt qui donne des r sultats comparables ceux de la figure 2 3 page 25 Comme d crit dans l article 39 la figure 1 2 montre bien la progression puis disparition de la jach re ainsi que l rosion des prairies 1 4 2 Visualisation de transitions entre cultures Pour obtenir un r sultat comparable celui de la figure2 1 page 23 il faut sp cifier un nouveau mod le HMM dit HMM ergodique avec tats de Dirac fau sens du maximum de vraisemblance Z caro
6. farming systems questions in which the activities of farmers and their changing location from the new picture is the focus point of problem solving 27 11 A number of new research tools such as remote sensing data and Geographical Information Systems are now available to address this type of research 8 In most cases farmers are seen to take into account the properties and layout of their land in deciding about the location of their cropping and grassland systems 53 This relationship between farmers and their territory could be an individual or a collective one 40 2 2 2 Land use is managed by farmers The land used by agriculture can be modeled as a complex and dynamic pattern of fields including tilled plots and pastures Sebillotte in the 1st European Society of Agronomy Congress defined the cropping system as a set of crop management procedures used on a homogeneously treated space inside a farm which can be a field or a part of field or several fields According to this definition a given cropping system is a component of a farming system and is identified characterized by the sequence of crops and corresponding technical operations 54 The cropping system is a tool to characterize land use on the tilled part of farms 55 However many farms have not only tilled crops but comprise 16 also pastures So if we want to reason at farm scale it is necessary to generalize the concept of cropping system by including grassl
7. il n est pas en Open Access Toutefois pour permettre une d monstration et v rifier la bonne installation du logi ciel ce fichier artificiel est fourni DLL est un dossier qui contient les DLL Dynamic Link Libraries Qt pour Windows Xp ou 7 Pour faire fonctionner CAROTTAGE il faut mod ifier la variable PATH dans le menu System de Windows pour ajouter le chemin d acc s ce dossier On peut aussi copier les fichiers du r pertoire DLL dans le r pertoire SrcQt debug 1 4 Exemple pour d buter Dans cette section nous allons ex cuter les diff rents outils de CAROTTAGE sur le fichier de d monstration fourni short example txt 2http www loria fr jfmari App Sattention l orthographe anglaise 1 4 1 Segmentation de la p riode d tude Il s agit d tudier la dynamique de l assolement de notre r gion Une premi re solution consiste d terminer autant d assolements qu il y a d ann es de collecte d occupation Ter Uti Une autre solution consiste se limiter un nombre limit de p riodes disons trois pour avoir une vue plus concise de l volution et de laisser les HMM effectuer la meilleure segmentation Nous utiliserons le fichier short example txt pour obtenir des r sultats comparables ceux de la publication 39 Dans cette fouille de donn es on s int resse aux observations form es d une seule occupation du sol Leurs d finitions sont regroup es dans le fichier terutil cfg
8. state at time t 4 There is a doubly stochastic process e the former is hidden from the observer and is defined on a set of states e the latter is visible It produces an observation the land use of a parcel at each time slot depending on the probability density function that is defined on the state in which the Markov chain stays at time t It is often said that the Markov chain governs the latter Thus a HMM2 is specified by e a set of N states called S s1 sy e a three dimensional matrix a over 5 Qijk Prob q Sk qi 1 8j 4 2 Si 2 1 Prob q Sk qt 1 Sj 2 Si U 3 with the constraints Sa aijk 1 Y i j 1 NP and where q is the current state at time t e a set of N discrete distributions b is the distribution of observations associated to the state s This distribution may be parametric non parametric or even given by an HMM The probability of the state sequence QT q1 q2 qr is defined as T Prob QT Lg Gq go Il Agi 2qt_10t 2 2 t 3 where Vj qj S Ig is the probability of state q and ay is the probability of the transition q q2 initialization of the model at times t 1 and t 2 20 Given a sequence of observations OT 0 0 07 the joint state output probability Prob QT OT is defined as T Prob Q Of Igba 01 q1q20q2 02 Il Qge_aqe 1qe0c 0t 2 3 t 3 The estimation of a HMM1 is usually done by the Baum We
9. these triples belong to the same beet wheat pea wheat four crops succession Clustering the districts of the Seine Basin The analysis of crop sequences and the determination of the main successions double triple or even quadruple successions as shown before are a basis for comparing and classifying agricultural territories The small districts and the sub watersheds of the Seine basin were compared and clustered thanks to statistical methods applied on the sets of crop triples that characterized each district or watershed Finally we built a district typology clustering the similar districts wrt the crop successions More precisely the analysis of Ter Uti data in the Seine Basin was performed following these steps 1 Determination of the main crop successions in each small district using CAROTTAGE model 2 as explained in Sections 2 4 2 and 2 4 2 The whole basin was characterized with 64 3 or 4 crops successions for 143 districts The crop successions were clustered within 6 main categories according to their agronomic function cereals break crops etc 2 Computation of the distribution of the crop successions in each small district using CAROTTAGE model 1 The districts were thus characterized for a period with sets of crop successions and their probabilities 3 Analysis of the table districts x probabilities of crop successions using the Principal Component Analysis method The projections of the districts on t
10. wheat 1 and wheat pea wheat 2 which have the thickest lines bottom of the diagram e the transitions beet pea 3 appear between 1992 and 1995 and then disappear e transitions like wheat barley 4 or barley beet 5 appear from 1996 actually they exist before 1996 but with a probability smaller than 1 5 One can also notice that the other crops like rapeseed maize potatoes or set aside are mainly followed or preceded with wheat Furthermore the transitions wheat wheat 6 seem to grow between 1996 and 1998 Three crop sequences in Saint Quentinois In order to better examine the crop transitions we transform the Ter Uti data and apply CAROTTAGE on tables representing couples triples or even quadruples of crops To minimize the data set we have to select the main 29 rotations based on our first analysis e g for crop triples wheat beet wheat wheat pea wheat etc Thus we obtain a second diagram where the states represent triples of crops which is more difficult to explain but confirms our first analysis Figure 2 8 pea w barley beet pea w wiw beet wiwiw beet w w barley beet w w barley beet beet w barley beet w beet ae TS rent AT wipea w ONE LE RS TX R DES DR HE DES beet w pea PSN DINNE INN 1992 94 1993 95 1994 96 1995 97 1996 98 1997 99 Minimal Probability for displaying 1 00 Figure 2 8 Transitions between triples of crop
11. 001 H Jonas Le principe de responsabilit Une thique pour la civilisation technologique Das Prinzip Verantwortung Editions du Cerf Paris 1990 1979 336 pages Michael Jordan and Zoubin Ghahramani Factorial Hidden Markov Models Machine Learning 29 2 3 245 273 November 1997 D L Karlen G E Varvel D G Bullock and R M Cruise Crop rotations for the 21st century Advances in Agronomy 1994 E F Lambin X Baulies N Bockstael G Fischer T Krug and et al Land use and land cover change LUCC implementation strategy IGBP Rep 48 IHDP Rep 10 Int Geosph Biosph Program Int Hum Dimens Glob Environ Change Program 1999 Stockholm Bonn E F Lambin H J Geist and E Lepers Dynamics of land use and land cover change in tropical regions Annual Review of Environment and Resources 28 205 241 2003 S Lardon J P Deffontaines J Baudry and M Benoit L espace est aussi ailleurs In J Brossier B Vissac and J L Le Moigne editors Mod lisation syst mique et syst me agraire D cision et organisation pages 321 337 INRA Paris 1990 F Le Ber and M Beno t Modelling the spatial organisation of land use in a farming territory Example of a village in the Plateau Lorrain Agronomie Agriculture and Environment 18 101 113 1998 40 39 40 Al 42 43 A4 F Le Ber M Benoit C Schott J F Mari and C Mignolet Studying Crop Sequences With Carro
12. 5th International Symposium of the association for farming systems research extension Pretoria pages 952 959 1998 J M Attonaty M H Chatelin and F Garcia Interactive simulation modelling in farm decision making Computers and Electronics in Agriculture 22 157 170 1999 C Aubry F Papy and A Capillon Modelling decision making processes for annual crop management Agricultural systems 56 45 65 1998 J K Baker Stochastic Modeling for Automatic Speech Understanding In D R Reddy editor Speech Recognition pages 521 542 Academic Press New York New York 1974 P Baudoux G Kazenwadel and R Doluschitz On farm effects and farmer attitudes towards agri environmental programs a case study in baden wtittemberg Etudes et Recherches sur les Syst mes Agraires et le D veloppement 1998 333 356 1998 Brossier J Dent B eds Gestion des exploitations et des ressources rurales Entreprendre n gocier valuer Farm and Rural Management New context new constraints new opportunities B Benmiloud and W Pieczynski Estimation des param tres dans les chaines de Markov cach es et segmentation d images Traitement du signal 12 5 433 454 1995 Marc Benoit Florence Le Ber and Jean Francois Mari Recherche des successions de cultures et de leurs volutions analyse par HMM des donn es Ter Uti en Lorraine Agreste Vision La statistique agricole 31 23 30 June 2001 37 8 11 12
13. CAROTTAGE Windows pour les donn es Ter Uti manuel d utilisation Mari Jean Fran ois Loria Inria Grand Est 615 rue du Jardin Botanique BP 101 F 54600 Villers les Nancy France February 21 2014 Chapter 1 CarottAge 1 1 Pr sentation de CarottAge CAROTTAGE est un r tro acronyme construit partir du mot carotte qui se traduit en markov en russe et du mot ge C est aussi un proc d d analyse de la constitution des sols Faire un carottage d un sol c est extraire par forage un cylindre repr sentatif des couches travers es afin d tudier leurs successions et les dater CAROTTAGE est le r sultat d un travail de fouille de donn es effectu par des agronomes de l Inra SAD ASTER Mirecourt et des informaticiens du projet ORPAILLEUR Loria et Inria Grand Est pour extraire des bases de donn es agricoles Ter Uti des informations sur les successions de cultures pratiqu es dans une r gion CAROTTAGE s appuie sur la th orie des chaines de Markov cach es HMM comme Hidden Markov Model pour permettre l analyse de succes sions d observations quelconques continues ou discr tes Ces mod les per mettent de repr senter des observations temporelles comme des successions d tats o les transitions entre tats d pendent suivant l ordre du mod le de l tat courant et des n tats voisins Le logiciel calcule et affiche un signal dont l analyse permet l extraction et la datation de r gularit s t
14. Le bon choix est le r pertoire Mod TerutiLucas Le choix du r pertoire des binaires Le bon choix et c est le seul est SrcPirenSpatial 1 5 2 Le sous menu donn es Cette version de CAROTTAGE traite des fichiers de donn es Ter Uti labor s partir de donn es fournies par le Service central de la statistique agricole Dans ce menu il faut ici pr ciser o se situe le fichier short example txt cf Fig 1 3 Ce menu permet aussi de se limiter une p riode d tude et d appliquer un filtre d extraction de points Ter Uti par exemple en pr cisant une liste de PRA ou de d partements Il faut pour cela avoir le fichier nouvelleFrance txt 1 5 3 Le sous menu configuration L enqu te Ter Uti fournit une classification tr s pr cise de l occupation du territoire La centaine d tiquettes diff rentes Ter Uti doit tre regroup e en un nombre bien inf rieur de classes d occupation du sol Commencer par choisir dans ce sous menu importer une configuration et choisir le fichier teruti1 cfg qui se trouve dans le dossier config Par une s rie de glisser ins rer on peut modifier ce regroupement Avant de sortir de ce sous menu valider la configuration ce qui cr era le fichiers des observations possibles M CarottAge Fichiers Donnees Configuration Hmm Apprentissage Diagramme de Markov Visualisation Help row header size Separateur de colonne Configuration Modele Diagramme visualisation
15. ands 28 So we propose to name Agricultural Land Management System ALMS the system of crop and grassland management procedures used on a portion of land which can be a field or a part of field including its boundaries or several fields According to this definition a given ALMS is a component of a farming system and is identified characterized by the choices of the rotation of crops or grassland uses the farmland structure and the location rules of the crop rotations and grassland uses This definition should be completed by including also common items such as hedges fences etc that are components of the landscape and play a role in farm management 15 For us the ALMS is the basic unit of landscape design at farm scale At a regional scale other land uses and actors outside farms should be taken into account forests waters wild areas to complete the ALMS according to the aims of the models biodiversity management water protection leisure as well as collective farmers organizations 17 18 30 58 57 2 2 3 A proposal of European notation for crop sequences identification As a tool of representation and understanding of the interactions between agriculture land and environment agricultural land use management could be used as well for research as for management and negotiations in agro environmental policies The main topic in this way should be focused on land use changes 35 36 Although the agricultural practice
16. aphic shown figure 2 5 six crops have been individualized the container state is denoted by The thickness of the lines represents the a posteriori probability of the transition between two crops cf Equation 2 10 in the Appendix Diagonal lines mean that a 25 crop is followed by another crop e g rapeseed denoted by colza to wheat denoted by ble while horizontal lines mean that a crop is followed AAA SY AN NN NN X mais colza orge ble ppp bois 1992 1993 1994 1995 1996 1997 1998 1999 Figure 2 5 Viewing the results of model 2 applied on land use data of the Lorraine Region years 1992 1999 These tables are very useful for seeing the evolution of the land use in a region and for comparing regions The models can be used on crop data but also on sequences of crops and allow to produce sets of tables showing the evolution or stability of land use Compared to HMM1 HMM2 have the capability to model the transitions between Dirac states over a longer period according to the farmer s practices three years compared to two years Furthermore these tables can be used as a support for field inquiries Finally CAROTTAGE allows the user to define various models according to the data format and his purpose as we see in the next part 2 4 Using CarottAge for finding out crop sequences 2 4 1 The data base The Ter Uti data are collected by the French agriculture administrat
17. asses to recognize As opposite to pattern recognition we do not have the knowledge of what to recognize but rather look for something regular to extract hence the name data mining Actually data mining can be defined as the use of algorithms to extract information and patterns from databases 23 22 These algorithms are able to search the data and attempt to fit a model to the data using some preference criteria Data mining is a part of knowledge discovery processes that include four other steps the selection of data the preprocessing of data the transformation of data and the interpretation of the data mining results 22 In the present work we specify one second order Hidden Markov Model HMM2 in order to model in a more simple way the unknown behavior of a crop sequence We rely on the assumption that the land use of a field at time t depends on the land use of the same field at time t 1 t 2 etc Each state of the HMM2 captures a stationary behavior and represents a class a crop or a cropping pattern where the observations are drawn with a known probability density function Furthermore we compute the a posteriori probabilities that the Markov chain goes through some states between certain time slots These a posteriori probabilities can be plot as a function of time and determine a fuzzy classification in the states space This classification can be interpreted by the agronomists wrt the evolution of crop patterns and cro
18. bability CAROTTAGE is written in C and runs under Unix and X11R6 systems It has been designed specifically for mining land use data based on HMM2 It is able to analyze temporal and spatial sequences of land use in a territory Several models are available we describe a few of them below The CAROTTAGE software is now used by agronomists and also by geneticians for mining genomic data 29 without any assistance of the designers http www loria fr jfmari App 21 The functionalities of CarottAge CAROTTAGE get as an input preprocessed or transformed discrete data represented within text files Data mining is performed in four steps 1 the editing of the initial model 2 the iterative ML estimation using the Baum Welch algorithm based on a corpus of sequences of observations 3 the display of the model s parameters 4 the display of the a posteriori transition probabilities The user has to write the initial model in two parts The first part specifies the model s topology by means of a list of transitions between the states together with their relative weights The second part defines the observations and gives the discrete probabilities over this set of observations An example of a text file specifying a simple three states left to right self loops HMM2 where the three states have a uniform distribution is described in Table 2 2 2 3 1 3 4 1 2 2 1 3 3 1 4 4 1 stl equiprobable state 2 eq
19. benoit schott mignolet mirecourt inra fr also in Ecological Modelling 191 1 170 185 Jan 2006 2 1 Introduction Sixty years after its launching through the Marshall Plan the European agriculture revolution is up again but with some strong contradictions water pollution landscape uniformization ethical crisis 26 These harmful side effects of agriculture could be aggravated if the evolution of agricultural practices continues following the current trends towards greater concentration intensification and technicality We focus on agricultural practices from their choice by farmers decisions to their effects as they continuously remodel the agricultural landscapes The 14 approach of farming systems as landscapes builders is a new one but its background is the vision of land as resource for agriculture 19 37 Agronomic measures specifically designed to maintain soil water and air quality are necessary including more severe regulations restricting intensification and the agricultural use of chemicals For instance keeping the nitrate content of drainage water to less than 50 mg l 1 requires not only an optimized and reduced application of fertilizers but also the planting of catch crops during the winter Parts of the hydrological basins in many areas should be withdrawn from arable cropping and turned into grasslands or forests several authors in 43 Preventing runoff erosion and the associated pollution o
20. du cas id al Lorsque le mod le ne correspond pas la r alit on assiste un ph nom ne de d rive dans lequel les tats de Dirac se peuplent d occupations majoritaires bien diff rentes de ce qui tait pr vu au d part Tout l art de la fouille de donn e par mod lisation stochastique consiste sp cifier des mod les qui convergeront 12 vers un mod le utile l extraction de connaissances 1 6 D veloppements futurs CAROTTAGE a donn naissance ARPENTAG 46 car la recherche des successions de cultures dans un territoire a vite fait appara tre le besoin de pouvoir les localiser et faire appara tre des quartiers culturaux comme l avait fait remarquer J P Deffontaines 20 Pour utiliser CAROTTAGE sur d autres jeux de donn es il est n cessaire de cr er un r pertoire SrcMonProjet pour y d river une classe partir de Corpus0 Cela n cessite un travail de programmation en C Les expli cations pour y arriver d passent le cadre de ce tutoriel et feront l objet du Manuel du programmeur 13 Chapter 2 Annexe Studying crop sequences with CARROTAGE a HMM based data mining software F Le Bert M Beno t C Schott J F Mari C Mignolet 1 ENGEES 1 quai Koch BP 1039 F 67070 Strasbourg CEDEX fleber engees u strasbg fr phone 33 388248230 fax 33 388248284 2 UMR 7503 LORIA BP 239 F 54506 Vandceuvre l s Nancy CEDEX jfmari loria fr 3 INRA SAD Domaine du Joly F 88500 Mirecourt
21. e conceived by farmers this has the advantages of corresponding to the planning structure of the farmer which reasons rotations over several years and to allow a stability of land use descriptions over years whereas crop by crop descriptions would vary each year However they lack the account of the logic behind the simple crop rotation description although some hints may be given such as maize for silage versus maize for sale which complete the raw fact description so these notations cannot yet be fully counted as ALMS nomenclature In the future our aim is to contribute to build a framework of farmer rules used to build rotations 13 The first work done by 3 shows the importance of delay between two crops sowing and harvesting dates machinery choices Examples of use of the proposed European cropping grassland management systems are given in table 2 1 For crops M wW wB ic means Maize winter Wheat winter Barley with intermediary crops in autumn after harvesting For grassland bC tPH2 means each year the uses are the same hC means mowing for hay making tPH2 means turning Pasture For Heifers 2 years old Table 2 1 Nomenclature of crops and grassland uses sequences Each crop name e g M wW or grassland use cluster e g hC tPH2 represents a year of the sequence 18 2 3 Temporal Data Mining with mm2 The purpose of pattern recognition is to specify as much models as there are cl
22. emporelles et spatiales Il est fourni sous forme d une bo te outils comportant plusieurs programmes ind pendants ainsi qu une application graphique qui permet de les encha ner d une fa on interactive La premi re publication majeure de CAROTTAGE se trouve dans la revue Ecological Modelling Studying crop sequences with CAROTTAGE a HMM based data mining software 39 dont le pre print est donn en annexe Sa lecture est vivement conseill e d sol avant toute exp rimentation 1 2 Pr sentation des donn es Ter Uti Notre ensemble de donn es est constitu de l enqu te Ter Uti qui est r alis e par un sondage a deux niveaux de granularit Un premier tirage r alis par VIGN consiste s lectionner des photos a riennes r guli rement r parties sur l ensemble du territoire m tropolitain Les photos repr sentent cha cune un carr de 2 km de c t et sont s par es en moyenne par 6 km Un deuxi me tirage r alis par les DRAF consiste superposer sur chaque photo une grille de 36 points Compte tenu de la distance entre les photos la repr sentativit d un point est proche de 100 hectares L ensemble de ces sites est visit annuellement par des enqu teurs qui rel vent les occupations des sites Pour plus de d tails sur la grille Ter Uti on peut se reporter 41 Outre la s quence temporelle des occupations de chaque point nous savons quelle PRA il appartient et nous connaissons s
23. es voisins c est a dire la disposition relative de chaque point et de chaque photo a rienne En revanche nous ignorons la localisation pr cise des points pour des raisons de secret statistique Les services de statistique de la DRAF ont r parti les occupations en diff rentes classes environ 80 qui vont de marais salants tangs d eau saum tre peupliers pars en passant par superficie en herbe faible productivit potentielle Certaines de ces classes ne sont pas ou peu pr sentes dans les r gions tudi es consid r es aussi avons nous restreint le nombre de classes 49 par regroupement ou suppression 7 nLig 112806 anneel 1992 anneen 2003 nAttr 1 indeter 95 isHeader 1 pt dep pra photo pti 92 93 Sa 02 1 2 2034 8885 1 27 28 2 2 2034 8885 2 27 33 3 2 2034 8885 3 27 40 Table 1 1 Chaque point est tiquet par son d partement dep et sa PRA petite r gion agricole Direction R gionale de l agriculture et de la For t 12km 2m 6km 4km re gt 6km Lango a Le principe du maillage de base du b Les 4 photos a riennes choisies territoire dans une maille 250m 1500m 250m gt lt gt 13 14 15 16 17 18 25 26 FA 28 29 30 31 3e 33 34 35 36 c La photo a rienne et sa grille 6x6 Figure 1 1 Description des do
24. f surface water especially by pesticides needs grassland strips ditches or other structures placed in suitable strategic locations in a catchment Again similar conclusions could be drawn about many other environmental targets such as biodiversity or landscape quality and accessibility 13 The farmer practices are the focus point of researchers who built tools to help their changes 9 In this paper we propose a methodological approach of farmer practices involved in the land designing through land uses and land pattern changes Actually our approach combines agronomic and artificial intelligence methods We rely on a land use data base that we explore with a data mining approach to find out spatial and temporal land patterns Mining sequential and spatial patterns is an active area of research in artificial intelligence One basic problem in analyzing a sequence of items is to find frequent episodes i e collections of events occurring frequently together We rely on new numerical algorithms based on high order stochastic models the second order hidden Markov models HMM2 capable to discover frequent sequences of events in temporal and spatial data These algorithms can extract spatial and temporal regularities that can be explained by human experts and may constitute elements of a knowledge discovery process 48 Thus agronomists and computer scientist have designed a data mining software named CAROTTAGE in order to extrac
25. ged in different farms and hence facilitates the extension of cropping system research to the territorial and multi year scales which are relevant to environmental questions The concept of agricultural land management system is a first step towards the precise description and classification of all types of land uses intervening in a region In order to understand and manage the evolution of landscapes it will be necessary to include non agricultural uses forests waters in marshes waters are subject to a particular type of collective management roads and roadsides etc With respect to this purpose CAROTTAGE has proven useful for exploring large land use data bases and for revealing the temporal and spatial organization of land use based on crop sequences 48 Furthermore CAROTTAGE can also be used to investigate and visualize the crop sequences of a few specific farms or of a small territory Besides the diagrams resulting from CAROTTAGE showing the main crop transitions are good graphical supports for discussing the evolution of land use For example they have been used during regional farm surveys to collect the knowledge of farmers and agricultural technicians about crop sequences Finally the results of our analysis can be linked to models of nitrate flow and used for the evaluation of water pollution risks in a watershed 52 To resume crop sequences are a pragmatic research object useful to explain land use changes and we pr
26. he fifteenth first eigenvectors were used to design a new table with 15 variables characterizing the districts 4 Clustering of the districts on the basis of this last table using the Hierarchical Ascendant Clustering method The districts were clustered within twenty classes which represented a good 31 segmentation according to agronomists expertise The map of the Seine Basin where the districts are colored according to this segmentation is displayed in figure 2 9 Temporary Grassland dominant tG tG tG Cereals C5 tcicis w rs w Maize dominant ZZ M W M W tG tG tG M M M M M M M M W M W M W JLi M W M W M M M M W B VIL M W M W W S W W RS W W W W M W W M M M W W W Vegetables Rape seed dominant z Rs W B M W B M W Rs W M M W Rs W Rs W M M M Rs W B S W Rs W Rs W W BA rs w s rs w w Undifferentiated Rs W B W P W W M W Field pea dominant 0 25 _50 100 150 Cie P W W Rs W W W W W P W Rs W W W B EN p w m w P W Rs W vs vu ve Sugar beet dominant ol P W Bt W aG aG aG W P W B Bt W B MM 8t w Bt w p w Bt w GUNN ve w at w w r w tn Po W Bt W P W Bt W Bt W Bt w Legend W wheat B barley Rs rape seed M maize Bt sugar beet x S sunflower P field pea Po potatoe tG temporary grassland aG Vegetables dominant p E 7 RE artificial grassland SESH Vegetables M W M W Po W Bt B Orl ans Figure 2 9 Map of the Se
27. hemselves and designed their own models as shown in Section 2 4 2 A first model can be used for the extraction of temporal segments in which the distribution of the land use categories is stationary To do so we have specified a HMM2 with n states with a left to right self loops topology see Figure 2 2 This means that we attempt to capture n periods of evolution in the land use dynamics where n is chosen according to the length of the period o 3 4 Figure 2 2 Model 1 the HMM2 performs a data segmentation in three periods in which the observations are supposed stationary This model is defined in Table 2 2 The results of this model are displayed within a table where the evolution of the cropping pattern of a region is visible Figure 2 3 Here we see that the pastures are dominant at the beginning of the period and then decrease and are replaced by wheat at the end of the period while the surface of rapeseed is continuously growing This table is actually a synthetic view of the eight years 1992 99 pointing out the stable patterns and the main transitions Another model has been designed for measuring the probability of a succession of three land use categories Actually we have defined a specific state called the Dirac state whose distribution is zero except on a particular land use category Therefore the transition probabilities between the Dirac states measure the probabilities between the land use categories d
28. ine Basin districts classified wrt their main crop sequences This map highlights the strong spatial structure of the distribution of crop sequences in the Seine Basin This structure is to be related to big geological forms as for example e The districts classified into Rs W B dominant are on the Jurassic calcareous plateaux east of the basin e The districts of the classes temporary grassland dominant correspond to the granitic mountains of Morvan south east of the basin 32 e The districts of the class Bt W Bt W dominant occupy the silty plateaux of Picardy north of the basin 2 5 Conclusion We claim that the concept of crop sequences is relevant and can be useful it will help research on agriculture environment relations by providing types of land use that convey the farmer s strategy independently of year to year changes that characterize crop rotations these types of land use are stable over several years and can be related on one side to field characteristics and constraints and on the other side to environmental effects It will facilitate discussions between farmers and other actors of rural territories by setting a common language and allowing an objective description of agricultural land use types This concept by considering the middle term strategy of the farmer frees itself from the infinite diversity of actual crop successions and facilitates the comparison between fields similarly mana
29. ion on the whole metropolitan territory They represent the land use of the country on a one year basis Two levels of resolution are achieved Figure 2 6 A first sample consists in selecting aerial photographs The French territory is segmented into 3820 meshes Each of the meshes contains four photographs that cover each one only a square of 2 km Secondly on each photography a 6 by 6 grid determines 36 sites that are inquired every year in June The land use category of these sites wheat corn potato forest rocks is logged in a matrix in which the rows are the sites of the country and the columns the time slots from 1992 to 2003 Finally one Ter Uti site represents roughly 100 hectares 42 2 4 2 Analyzing crop sequences in the Seine Basin For thirty or forty years the increasing human activities domestic industrial agricultural have gradually degraded the hydro system of the Seine river regarding water quality and biological population 50 The nitrate contamination of groundwater and surface water is mainly caused by the evolution of agricultural activities and related to their nature and to their organization inside the river watershed The INRA team in Mirecourt is member of an interdisciplinary research program which aims to develop a tool for forecasting water quality in the Seine river watershed based on assumptions upon agricultural changes Thus the INRA team analyses the agricultural activities in the watershed
30. lch algorithm which is related to the EM algorithm 21 We have shown that a HMM2 can be estimated following the same way 45 The estimation is an iterative process starting with an initial model and a corpus of sequences of observations that the HMM2 must fit Usually the initial model has equiprobable transition probabilities and an uniform distribution in each state At each step the Baum Welch algorithm determines a new model in which the likelihood of the sequences of observation increases Hence this estimation process converges to a local maximum according to the maximum likelihood ML estimation criteria 21 47 To assess the final model we use the Kullback Leibler distance between the distributions associated to the states 56 Two states that are too close are merged and the resulting model is re trained Intuitively the Baum Welch algorithm counts the number of occurrences of each transition between the states and the number of occurrences of each observation in a given state in the training corpus Each count is weighted by the probability of the alignment between the states and the observations cf Equation 2 3 The principles of this algorithm are detailed in the appendix 2 3 2 CarottAge CAROTTAGE is a free software under a Gnu Public License that takes as input an array of discrete data the rows represent the spatial sites and the columns the time slots and builds a partition together with its a posteriori pro
31. luwer Academic Publishers January 2001 Jean Francois Mari and Florence Le Ber Temporal and spatial data mining with second order hidden markov models In Mohamed Nadif Amedeo Napoli Eric San Juan and Alain Sigayret editors Fourth International Conference on Knowledge Discovery and Discrete Mathematics Journ es de l informatique Messine JIM 2003 Metz France pages 247 254 IUT de Metz LITA INRIA Sep 2003 A1 49 50 51 52 53 54 56 57 J F Mari F Le Ber and M Beno t Fouille de donn es agricoles par mod les de markov cach s In C 2000 Journ es Francophones d Ing nierie des Connaissances Toulouse pages 197 205 AFIA ERSS IRIT GRACQ 2000 M Meybeck G De Marsilly and E Fustec La Seine en son bassin fonctionnement d un syst me fluvial anthropis Elsevier 1998 750 pages B J Miflin Sugar beet production strategies for the future In Proceedings of the 60th IIRB Congress pages 253 262 IIRB Brussels 1997 C Mignolet C Schott and M Beno t Spatial dynamics of agricultural practices on a basin territory a retrospective study to implement models simulating nitrate flow The case of the Seine basin Agronomie 24 2004 219 236 2004 P Morlon and M Beno t tude m thodologique d un parcellaire d exploitation agricole en tant que syst me Agronomie 6 499 508 1990 M Sebillotte Some concepts for analysing farming and cropping sy
32. n cliquant sur chaque tat on peut choisir les occu pations pour les transformer en tat de Dirac Dans notre cas il faut choisir une tat bl un tat orge un tat colza un tat ma s un tat prairies et for ts en s lectionnant dans la liste toutes les occupations que l on souhaite capter par cet tat Le dernier tat reste quiprobable Une fois la de scription sp cifi l option cr er le HMM cr e la forme interne du HMM La description est un fichier texte cf Tab 1 2 alors que le HMM a un format interne binaire stock dans un fichier d extension mod 10 1 5 5 Le sous menu Apprentissage C est le moins fourni de tous mais celui qui en fait le plus Choisir un nombre d it rations gal aux nombres d tats sauf si vous savez ce que vous faites et lancer apprentissage par la commande fwtInra 1 5 6 Le sous menu diagramme Ce sous menu permet la visualisation des diagrammes de Markov cf Fig 1 6 et leurs sauvegardes dans diff rents formats E CarottAge Fichiers Donnees Configuration Hmm Apprentissage Diagramme de Markov Visualisation Help Choisir Probabilite Quadrillage Annee de debut Legende Lire Hmm gph Display Print Clear Screen Saveas v Donnees Configuration Modele Diagramme Visualisation 0 7 bois AE RS AAA DPO 0 9 prairies _pp MX XXNX 0 VUUVUYVUUCUUCLONE 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 Figure 1
33. nn es Ter Uti 3820 mailles quadrillent la France toutes ne sont pas repr sent es 4 photos a riennes sont choisies dans une maille une grille 6x6 d termine 36 sites 1 3 Installation On trouve CAROTTAGE pour traiter les donn es Ter Uti sur le site du Loria 2 sous la forme d une archive carottage windows teruti V1 zip Carot tage pour windows et donn es Ter Uti Cette version est param tr e pour traiter des donn es Ter Uti Plusieurs dossiers sont fournis SrcQt contient l ex cutable graphique carrotage exe dans le sous r pertoire SrcQt debug SrcPirenSpatial contient les binaires ex cutables compil s correspondant a tous les outils n cessaires pour traiter les donn es Ter Uti pour la France enti re config contient les fichiers de configuration Ter Uti Il s agit de fichiers donnant la d finition de la classification Ter Uti bl orge ainsi que les regroupements que nous avons op r s comme bois et eau qui regroupent toutes les superficies en bois et eaux respectivement Mod contient les fichiers de l espace de travail description de mod les mod les initiaux et finaux Corpus contient le fichier de donn es Ter Uti short example txt Le fichier de donn es Ter Uti NouvelleFrance txt construit partir des donn es Ter Uti fournies par le Service de la Statistique et de la Prospective SSP du Minist re en charge de l agriculture n est pas inclus dans ce dossier car
34. o Crop science scientific and ethical challenge to meet human need In 3rd International crop science congress 17 22 august 2000 Hambourg Germany 2000 11 pages F Gaury Syst mes de culture et teneurs en nitrates des eaux souterraines Dynamique pass e et actuelle en r gion de polyculture levage sur le p rim tre d un g te hydromin ral Doctorat de l Ecole Nationale Sup rieure Agronomique de Rennes 1992 R Gras M Beno t J P Deffontaines M Duru M Lafarge A Langlet and P L Osty Le fait technique en agronomie Activit agricole concepts et m thodes d tude INRA L Harmattan 1989 160 pages S bastien Hergalant Bertrand Aigle Bernard Decaris Jean Francois Mari and Pierre Leblond HMM an Efficient Way to Detect Transcriptional Promoters in Bacterial Genomes In Furopean 39 30 33 34 39 36 37 uu Conference on Computational Biology ECCB 20038 Paris France pages 417 419 Sep 2003 poster in conjonction with the french national conference on Bioinformatics JOBIM 2003 S Herrmann S Dabbert and H G Schwarz von Raumer Ecological threshold values as indicators for biodiversity economic and ecological consequences Agriculture Ecosystems and Environment pages 493 50 2003 J Adibi and W M Shen Self Similar Layered Hidden Markov Model In 5th European Conference on Principles of Knowledge Discovery in Databases Freiburg Germany September 2
35. on s gt Sj gt Sk between t 1 and t 1 during the emission of the observation sequence mi j k Prob q_ 1 Si qt Sj Qi sk 0T 013 0 OT 2 8 We deduce for all t 2 T 1 and for all i j k 1 N m i j k a i j aijrbe 0t41 bt1 J k Prob Of 01 0r 2 9 As in the first order we define Prob 1 Si qt s OF 01 07 amp i j as the a posteriori probability that the stochastic process accomplishes the transition s sj between t 1 and t assuming the whole sequence We obtain for all t 2 T 1 and for all i j 1 N N amp li j X mA 2 10 k 1 When the training corpus is a set of sequences we sum 7 7 over this set and plot this value as a function of t i and j are dropped in the Y axis This illustrates the behavior of the stochastic process between states s and sj at time t see Figure 2 5 The second order ML estimate of ajjx is given by the equation Tik gt mli j k XC m i j k 2 11 t k t If N is the number of states and T the sequence length the Baum Welch algorithm has a complexity of N x T for a HMM2 Interested readers may refer to 21 47 to find more specific details of the implementation of this algorithm 36 Bibliography 1 2 F Affholder P Bonnal D Jourdain and E Scopel Small scale farming diversity and bioeconomic variability a modelling approach In Proceedings of the 1
36. opose to apply our analysis method and the CAROTTAGE software to understand the recent changes and to 33 forecast the future new land uses 25 10 So logically our work will take place in the international project LUCC 35 To go further we have to enlighten the farmers about the links between their objectives their practices and the consequences of their practices 28 A possible approach is to test different scenarios for the actors Two types of scenarios may be developed based on the following argumentation What if and How to Research methods to address these two types of scenarios taking into account the analysis of farmer practices and modeling of decision making are to be developed 1 2 The model building process itself can serve as a tool to construct and discuss scenarios with the actors 16 Two main model building procedures are used mathematical ones involving methods used in landscape ecology and linear programming and graphic ones We shall elaborate on the second procedure since the first one is well known For example one research approach developed by geographers is to define a dictionary of spatial graphic symbols or chorems 14 Using this form of qualitative modeling proves most useful in discussions with a wide number of people and enables us to build models of farmer practices in their spatial dimension 20 A potential further development in this direction is the use of 3D visuali
37. p sequences 2 3 1 HMM2 definition and automatic estimation The second order Hidden Markov Models are based on the probabilities and statistics theories They are implemented with unsupervised training algorithms like the EM algorithm 21 that allow to estimate a model parameters from a corpus of observations and an initial model The resulting model is capable to segment each sequence in stationary and transient parts and to build up a classification of the data together with the a posteriori probability of this classification This characteristic makes the HMM2 s appropriate to discover temporal and spatial regularities as it is shown in various areas e g 6 12 24 31 Furthermore the very success of the HMMs is based on their robustness even when the considered data do not suit a given HMM its use can give interesting results In a HMM2 the underlying state sequence is a second order Markov chain Therefore the probability of a transition between two states at time t 19 depends on the states in which the process was at time t 1 andt 2 A Markov chain is defined over a set of states the crops in a field or more generally the land use categories in a place that are unambiguously observed The Markov chain specifies only one stochastic process whereas in a HMM the observation of a land use category is not uniquely associated to a state but is rather a random variable whose conditional density depends on the current
38. posteriori probabilities of transitions between the states Appuyer sur une touche pour quitter Ppp colza Contenu de l tat 2 00 Sols_a_couverture_boisee 0 718153 01 Sols_artificialises_non_batis 0 070581 02 autres_sols_ni_alteres_ni_batis 0 061851 03 prairies_temporaires 0 025860 04 Eaux_permanentes_et_zones_humides 0 025758 05 Sols_batis 0 022116 06 pre vergers 0 020906 07 jacheres 0 018205 08 potagers 0 010541 09 superficies_en_herbe 0 009016 10 zones_interdites 0 004016 136 D3 11 prairies_artificielles 0 003923 12 melanges_des_6_especes 0 001782 13 avoine 0 001345 1998 1999 fichage 1 00 Figure 2 1 Displaying the results of CAROTTAGE the user can see the dis tributions associated to the states table and the a posteriori probabilities of transitions between states diagonal and horizontal lines 23 Models for mining land use data The role of the user is obviously crucial it has to preprocess and transform the data to define the initial model and to interpret the data mining results Furthermore these actions can be combined in a knowledge discovery process where the data can be transformed in several ways and mined with various models Actually at the beginning of our work the models were defined and experimented by the users agronomists and the computer scientists together 49 Then the agronomists used CAROTTAGE by t
39. rward probability is defined for all j k L N as atlj k Prob q 1 Sj qt sk Of 0 nor 2 4 This value represents the probability of starting from the initial state s1 and ending with the transition sj Sp at time t and generating output 01 04 using all possible state sequences in between The Markov assumption allows the recursive computation of the forward probability for t 3 T as follows N alj k gt at i j Qijk br os 2 5 i 1 Without any loss of generality we can suppose that sy is the only final state then the probability that the model generates the sequence OT 0 o7 is Prob OT o1 o7 gt ar j N Another useful quantity is the backward function 7 7 defined as the probability of the partial observation sequence from t 1 to T given the transition s s between times t 1 and t It can be expressed for all t 2 T 1 and for all i j 1 N by bili j Prob OZ Ot 1 or qt 1 Si qt sj 2 6 The Markov assumption allows also the recursive computation of the backward probability as 1 Initialization Br i J 1 V i j 1 NP 2 Recursion for T 1 gt t gt 1 N Bilt j XO BeyrG amp aije de Or41 Wi j L NP 2 7 R Given an observation sequence 01 07 we define for all t 2 7 1 and for all i j k 1 N the value i j k as the probability of the 39 transiti
40. s 1992 1997 in the district of Saint Quentinois Only the transitions whose probability is greater than 1 0 are displayed The question mark denotes the container state For example if we follow the sequence of the crop triples starting from beet wheat beet in 1992 94 the main transition 1 leads to the wheat beet wheat triple From this last triple there are several possibilities a first one goes towards the triple beet wheat wheat 2 a second one towards beet wheat barley 3 a third one which has the greatest probability towards beet wheat beet 4 and finally a fourth one towards beet wheat pea 5 Knowing that two triples are connected when they share two crops we can synthesize the last transitions in the following way 2 beet wheat beet wheat wheat 3 beet wheat beet wheat barley 4 beet wheat beet wheat beet 5 beet wheat beet wheat pea Furthermore we notice a repeated pattern in this diagram that looks like a 30 chain link this pattern is composed with the repeated transitions between the triples wheat beet wheat and beet wheat beet and reveals the existence of the quadruple succession beet wheat beet wheat Another pattern made of oblique lines can be found in this diagram for example the line starting from the beet wheat beet triple in 1992 94 connects to wheat pea wheat then pea wheat beet wheat beet wheat and finally to beet wheat beet or again to beet wheat pea This connected sequence proves that all
41. s we are familiar with are far from covering the whole range of existing systems we shall propose a method for establishing a nomenclature of ALMS The origin of these proposals lies in a number of monographs done for a large diversities of farms in a European research project 5 This first large range of landscape building monographs meets the work described in 53 So we propose a common notation of land use descriptions Table 2 1 with two characteristics i description of the land uses as they are described managed and decided by the actors ii account of time scales as first organizational factor All over Europe and each year the farmers have to allocate their crops and grassland uses in their territory This allocation is an important part of farmer decision that we have to model 16 This annual adjustment IRegional Guidelines to Support Sustainable Land Use by EC Agri environmental Pro grams EAP AIR 3 CT94 1296 17 between chosen crops and field plots results in different perennial rotations of crops and grassland use types 34 Examples are e in Denmark maize maize winter wheat barley e in south west France without irrigation sunflower winter wheat barley e in the East region of France oil rapes winter wheat e in the plain of Rhine in Vorarlberg Austria maize maize temporary grassland for mowing 3 years These notations describe yearly sequence of crops or pasture uses as they ar
42. sont diff renci s se fait en plusieurs temps 1 sp cification d un mod le lin aire avec le m me nombre d tats 9 savoir 6 tats bl orge colza mais prairies for ts ainsi que l tat qui joue le r le de container et qui capturera toutes les exceptions cf Tab 1 2 2 transformation de ce HMM lin aire en ergodique par la commande lin_to_ergo 3 estimation par la commande fwtInra 4 visualisation du diagramme de Markov de la figure 2 3 page 25 Toutes ces tapes sont regroup es dans le fichier do markov bat cf Tab 1 3 NOOR W N NN O OP w ND h NO O1 W D O U1 W N bhhbhhhhhhhhhh 1 oa M M i ble orge colza mais prairie bois equiprobable Table 1 2 Description du HMM lin aire 6 tats dont 5 tats de Dirac fichier bocm 1st rem rem rem rem rem rem rem set rem rem commandes pour realiser les exemples de Ecological Modeling Studying Crop Sequence whith Carottage Leber Benoit Schott Mari Mignolet 2006 cultures simples teruti fichier terutil cfg executer dans Mod TerutiLucas CORPUS Corpus TerutiLucas short example txt remplacer terutii cfg par le fichier de configuration correspondant creation du Hmm lineaire a 6 etats start W SrcPirenSpatialWindows ter2indice tempo exe t config terutii cfg CORPUS o bin terutil 1st start W SrcPirenSpatialWindows editmodel exe
43. stems and for understanding their different effects In A Scaife editor Proceedings of the first Congress of European Society of Agronomy Colmar volume 5 pages 1 16 European Society of Agronomy 1990 M Sebillotte Syst me de culture un concept op ratoire pour les agronomes In L Combe and D Picard editors Les syst mes de culture pages 165 196 INRA ditions Paris 1990 J T Tou and R Gonzales Pattern Recognition Principles Addison Wesley 1974 P M van Dijk F J P M Kwaad and M Klapwijk Retention of water and sediment by grass strips Hydrological Processes 10 8 1069 1080 1996 P M Van Dijk M Van der Zijp and F J P M Kwaad Soil erodibility parameters under various cropping systems of maize Hydrological Processes 10 8 1061 1067 1996 42 59 P Vereijken A methodic way to more sustainable farming systems Netherlands Journal of Agricultural Science 40 209 223 1992 43
44. t config terutii cfg d bocm lst i bin terutil 1lst o lin bocm mod rem transformation en ergodique start W SrcPirenSpatialWindows lin_to_ergo exe t config terutil cfg lin bocm mod ergo bocm mod start W SrcPirenSpatialWindows fwtInra exe t config terutii cfg n 6 ergo bocm mod o ergo bocm modi CORPUS start W SrcPirenSpatialWindows gviewmod exe t config terutil cfg ergo bocm modi o ergo bocm txt m 10 start W SrcPirenSpatialWindows fwtInra exe t config terutii cfg n 1 x 2 ergo bocm modi o ergo bocm gph CORPUS start W GviewGraph2_Qt debug GviewGraph exe config terutii cfg ergo bocm gph 0 01 1991 2003 Table 1 3 do_markov bat fichier de commandes pour cr er le diagramme de Markov de la figure 2 1 1 5 Utilisation de l interface graphique CarottAge L archive contient une application graphique qui permet aussi d enchainer manuellement ces tapes en dispensant l utilisateur de l criture des fichiers de commandes Les r sultats sont les m mes dans les deux modes de fonc tionnement fichier bat ou interface graphique 1 5 1 Premi re utilisation A la premi re utilisation CAROTTAGE demande de choisir deux r pertoires un r pertoire de travail qui contiendra les fichiers de donn es ainsi qu un r pertoire de binaires Il est possible de revenir sur ces choix gr ce l option Fichier Le choix du r pertoire de travail
45. t crop sequences and patterns from land use data bases This software allows the user to specify the architecture of the Markov model according to the data and his objectives Displaying tools have also been defined CAROTTAGE is used in several research projects e g agronomists try to find out crop sequences in order to model nitrate loss due to agricultural activities The paper is organized as follows Part one is about the relationship between land and farmer practices and the modeling of crop rotations 15 Part two is about HMM2 and the CAROTTAGE software Part three presents some results obtained by CAROTTAGE on a French data base and their analysis Then we conclude and propose some perspectives 2 2 An agronomic question 2 2 1 The relationships between land and agriculture We want to focus on the mutual relationship between land and farmer practices on the one hand the current state of the land is a result of farming practices and changes in landscapes could not be decided without farmers participation but on the other hand the choice and location of cropping and grassland systems by farmers all over the world takes into account their own land characteristics 37 38 This management of land by farmers is a part of the global technical management building agriculture 9 and is a factor of farm economical effectiveness 2 The future of European land is based on this management 25 Environmental issues may be converted into
46. tAge a HMM Based Data Mining Software Ecological Modelling 191 1 170 185 Jan 2006 P Y Le Gal and F Papy Coordination processes in a collectively managed cropping system double cropping of irrigated rice in senegal Agricultural systems 57 135 159 1998 M Ledoux and S Thomas De la photographie a rienne la production de bl Agreste la statistique agricole 5 juillet 1992 M Ledoux and S Thomas De la photographie a rienne la production de bl AGRESTE la statistique agricole 5 1992 G Lemaire and B Nicolardot editors Ma trise de l azote dans les agrosyst mes INRA Editions Paris 1997 333 pages A Lovett S Herrmann K Appleton and T Winter Landscape modelling and visualisation for environmental planning in intensive agricultural areas In E Buhmann and S Ervin editors Trends in Landscape Modeling pages 114 122 Wichmann Heidelberg 2003 J F Mari J P Haton and A Kriouile Automatic Word Recognition Based on Second Order Hidden Markov Models IEEE Transactions on Speech and Audio Processing 5 22 25 January 1997 Jean Francois Mari El Ghali Lazrak and Marc Benoit Time space stochastic modelling of agricultural landscapes for environmental issues Environmental modelling amp software 46 219 227 August 2013 http hal inria fr hal 00807178 PDF arpentage_hal pdf Jean Francois Mari and Ren Schott Probabilistic and Statistical Methods in Computer Science K
47. ttage trunk Reco S pat ioTemp Mod Terut iLucas gt more lin3 txt Weight 1 686688612 3494974673 8 349497 bois 0 511708038786 8 162206 6182759255 8 186572 6820589378 8 063783 738529522 6564766 7989081671 80 0523706 8327322606 6 641832 8744441085 0 041711 92084086493 0 03392645 9275236637 6 619115 1 990000013 3222218454 322222 6 216867 prairies_pp 8 123612 ble 8756227 colza 60 0545815 orge 8 044945 Sols_art_NB 0416224 i 08 0244413 08 0224466 9408122264 80 0150517 1 000000013 3156380653 8 315638 4738562353 6 158212 prairies_pp 6163401902 8 14249 ble 7013884336 08 0850482 orge 8 784973003 08 0835846 colza 8340020031 8 849829 Sols_art_NB 8822738528 6 8482718 i 0 928042506199 08 0219768 9238443779 6 6187938 G 9389689844 6159246 pre vergers Weight Weight OOOO OOOO C6 eee 6 L 2 3 4 5 6 7 8 9 g 6 1 2 3 4 5 6 7 8 9 g 6 1 2 3 4 5 6 7 8 9 iZ carottage trunk Reco S pat iolemp Mod Terut iLucas gt Figure 1 2 Visualisation des 3 pdf de 1in3 txt L adjectif ergodique signifie ici que toutes les transitions entre tats sont possibles Le terme tat de Dirac a t emprunt a la th orie des dis tributions Il signifie que la densit de probabilit associ e cet tat a la forme d une impulsion de Dirac un pour une occupation z ro ailleurs La construction du mod le ergodique dans lequel les tats associ s au bl orge et colza
48. tural decision support systems Agricultural systems 52 355 381 1996 S Dabbert S Herrmann G Kaule and M Sommer editors Landschafts modellierung f r die Umweltplanung Springer Verlag 1999 260 pages S Dabbert S Herrmann T Vogel T Winter and H Schuster Socio economic analysis and modelling of agricultural water demands and land use In German Programme on Global Change in 38 24 25 26 28 Ch Hydrological Cycle Status Report 2002 Phase I 2000 2003 2002 55 pages C T de Wit Resource use efficiency in agriculture Agricultural Systems 40 125 151 1992 J P Deffontaines J P Cheylan S Lardon and H Th ry Managing rural areas From pratices to model In J Brossier L de Bonneval and E Landais editors Systems studies in agriculture and rural development Science Update pages 383 392 INRA Paris 1994 A P Dempster N M Laird and D B Rubin Maximum Likelihood From Incomplete Data Via The EM Algorithm Journal of Royal Statistic Society B methodological 39 1 38 1977 M Dunham Data Mining Prentice Hall 2003 U Fayard G Piatesky Shapiro P Smyth and R Uthurusamy editors Advances in Knowledge Discovery in Data Minig AAAI MIT Press 1996 Shai Fine Yoram Singer and Naftali Tishby The Hierarchical Hidden Markov Model Analysis and Applications Machine Learning 32 41 62 1998 L Fresco editor Future of the Land Wageningen 1993 L Fresc
49. uiprobable state 3 equiprobable state 4 Table 2 2 Initial model 1in3 mod the first lines describe the transitions a line is structured like lt origin gt lt extremity gt lt weight gt the last lines describe the distributions associated to the states Here the hidden states are called 2 3 and 4 The distributions are uniform Non uniform distributions can be also defined Then the state is described with a list of observations and their probabilities as follows 1 wheat state n which means that the state n contains only wheat and that the probability of the other observations is null CAROTTAGE provides a program that builds a file containing the HMM2 according to the text file used as input 22 The model is then estimated on a corpus of sequences represented by a matrix of observations A typical command line is estimate n 3 lin3 mod o lin3 modi lorraine xls The input file 1in3 mod cf Table 2 2 is estimated using the corpus specified by the file lorraine xls Three iterations are performed The resulting model is stored in the output file 1in3 mod1 Actually this file records the a posteriori transition probabilities see Equation 2 10 in the Appendix between the states and the distributions of observations associated to the states A specific program has been developed for displaying the results of the model estimation Figure 2 1 It displays both the model s parameters especially the distributions and the a
50. uring a three years period Figure 2 4 shows the topology of a HMM2 that has two kinds of states Dirac states associated to the most 3The number of steps is constrained by the memory of the HMM2 2 24 sea EE state pastures 0 31 pastures 0 29 wheat 0 29 wheat 0 22 wheat 0 26 pastures 0 27 barley 0 16 rapeseed 0 14 rapeseed 0 17 rapeseed 0 12 barley 0 11 barley 0 12 maize 0 07 maize 0 08 maize 0 06 set aside 0 05 set aside 0 05 orchard 0 02 Figure 2 3 Viewing the results of model 1 applied on land use data of the Lorraine Region years 1992 1999 frequent land use categories wheat maize barley and container states associated to uniform distributions over the set of observations The estimation process usually empties the container state of the land use categories associated with Dirac states Figure 2 4 Model 2 the states denoted 2 3 and 4 are associated to a distribution of land use categories as opposite to the states denoted with a specific land use category The number of columns determines the num ber of time intervals periods A connection without arrow means a two directional connection As results the user obtains a graphic showing the main transitions between Dirac and container states i e the crop sequences in a region Figure 2 5 The user can choose the resolution level and see all transitions or only the main transitions In the gr
51. zation tools to facilitate the understanding of the land use and landscape changes see 44 for an example To end with an ethical posture 32 we propose a new researcher behavior investigating this type of issue we must not set out from the assumption that a farmer has voluntarily deteriorated the landscape parameter that is being investigated This corresponds to the development of a decision agriculture 51 that is increasingly knowledge based and increasingly rooted in the information and communication sciences and technologies and to a sustainability trend with a new weight of land capabilities 59 33 We agree with 13 This does not however mean a technology driven process of innovation but on the contrary increased feedback of action and decision into the design of innovation mainly on land design management innovation 2 6 Appendix The Baum Welch Algorithm The Baum Welch or Forward Backward algorithm implements a HMM2 s estimation following the maximum likelihood estimation criteria Since many state sequences may generate a given output sequence the Land Use and Cover Changes 34 probability that a model generates a sequence OT 01 0r is given by the sum of the joint probabilities given in equation 2 3 section 2 3 1 over all state sequences 1 e the marginal density of output sequences To avoid combinatorial explosion a recursive computation can be used to evaluate the above sum The fo

Download Pdf Manuals

Related Search

Related Contents

Nexio AMP Protocol 7.5 Developers Guide  Homologation  User Manual  COL`EXTREME  

Copyright © All rights reserved. Failed to retrieve file