Agricultural landscapes are composed of many land management units. Involved stakeholders or specific research foci can define these units differently (Straume, 2014; Zanten et al., 2013); therefore, their study requires innovative approaches able to address temporal and spatial dynamics using multiple data sources (Brown et al., 2013). Methods to do so, in the literature, differ mainly in disciplinary backgrounds and study targets (e.g. environmental protection, conservation of cultural features). In this context, agronomy appears to have a marginal role because of relatively little interest in spatially explicit and context-related issues in agriculture. Accordingly, the emerging landscape agronomy field claims for to increase understanding of interactions between farming practices and natural resources at the landscape level (Benoît, Rizzo et al., 2012). We aimed to develop a method able to handle heterogeneous spatial data when defining land management units. We tested a stochastic data mining method originally developed for temporal and spatial modelling of agricultural land uses (Mari, Lazrak, & Benoît, 2013). We stressed the Markov random field (MRF) assumption of this method by assuming that characteristics of a spatial unit depend on characteristics of neighbouring units. The study was carried out on a Mediterranean terraced olive grove farming system (62 km 2 , Monte Pisano, central Italy). Different sets and classifications of variables were tested based on natural and management issues. Finally, the landscape was segmented using six variables: geology, aspect, morphology, land cover, terrace type and proximity to roads. The layers were sampled on a regular point grid, and then the MRF was approximated to a hidden Markov model using a space-filling curve. Results consisted of a set of maps of agro-environmental land management units and a hierarchy of related landscape characteristics. This exploratory method can improve landscape research by providing a rapid assessment of heterogeneous data in a spatially explicit way.
_______________
suggested citation: Rizzo D, Mari JF, Marraccini E, Lazrak EG (2014) Agricultural landscape segmentation: a stochastic method to map heterogeneous variables. 1st IALE-Europe Thematic Workshop: Advances in Spatial Typologies: How to move from concepts to practice? Lisbon (Portugal). http://bit.ly/1Ag59mV
Hubble Asteroid Hunter III. Physical properties of newly found asteroids
Agricultural landscape segmentation: a stochastic method to map heterogeneous variables
1. AGRICULTURAL LANDSCAPE SEGMENTATION
a stochastic method to map heterogeneous variables
Agricultural landscape segmentation • Rizzo et al. • Symposium 1
2. Researchgroup
RIZZO Davide
INRA SAD-ASTER
Mirecourt(France)
MARI Jean-François
Universitéde Lorraine, LORIA & CNRS
UMR 7503
Vandoeuvre-lès-Nancy F-54506, France
INRIA
Villers-lès-Nancy, F-54600 France
MARRACCINI Elisa
ScuolaSuperioreSant’Anna
Institute of Life Sciences Pisa (Italy)
LAZRAK El Ghali
INRA SAD-ASTER
Mirecourt(France)
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 2
3. Introduction#1 | Agriculturallandscapemanagement
Agricultural landscapes are composed of many land management units.
Involved stakeholders or specific research foci can define these units differently
Innovative approaches to address temporal and spatial dynamics using multiple data sources
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 3
biblioat http://bit.ly/IALE2014
Straume, 2014 Zantenet al., 2013
Brown et al., 2013Rizzo et al., 2013
4. Introduction#2| Landscapeagronomy
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 4
Agricultural land uses:
randomly distributed among fields
managed by farmers ?
Understanding how
spatial organization & temporal evolution
of farming practices
reveal logical processes and driving forces
Benoît, Rizzo et al., 2012
a major challenge for landscape agronomists
biblioat http://bit.ly/IALE2014
5. Aim| context-wiselandscapesegmentation
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 6
Arpentageis a HMM based software specifically developed to perform (un)supervised segmentations of agricultural landscapes
Developed to map crop sequences as support to evaluate agricultural impact on water quality and on biodiversity
Mari et al. 2013
biblioat http://bit.ly/IALE2014
Test this method in new contexts with an explicit landscape agronomic approach
Profileof heterogeneousspatialdata insteadof (crop) time sequences
Focus on abandonedareasin a Mediterraneanterracedlandscapedominatedby olive groves
Mignolet2008
Lazraket al. 2009
Temporospatialmethod…
…to maplandscaperegularities
HMM= Hidden Markov Model
6. Methods#1 | Landscapesegmentation
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 7
Mari et al. 2013, EMS • Lazraket al. 2009& Schaller et al. 2012, Lands Ecol
biblioat http://bit.ly/IALE2014
Space dominant
sequenceof images
Software for LUCC analysis generally implement unsupervised clustering on spatial entities based on space-dominant attributes
Data mining approaches involving clustering of sequences allowed knowledge extraction to get a landscape description in terms of land-use patterns
Time dominant
image of sequences
7. IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 8
temporal segmentation
Linear
Ergodic*
spatial segmentation (andtransition analysis)
*all states are inter connected
Markov chain
land use of a given field in a year depends only on the land-use of the recent previous years in the same field.
Markov random field
land use of a given field depends only on the land-use of the neighboring fields
probably the 1stsoftware in the agronomic area implementing a time-dominant model and processing time-space data.
It relies on a stochastic principle to model time space landscape regularities on 2 Markovianassumptions
free download: http://www.loria.fr/~jfmari/App
Methods #2 | ArpentAge
French acronym for: Landscape Regularities Analysis: Environment, Territory, Agronomy
Arpenteris a French verb meaning to survey
8. Methods #3 | Approximating a MRF by a HMM
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 9
the Hilbert-Peanofractalcurve isusedto approximatethe MarkovRamdomField (MRF) assumptionthrougha HiddenMarkovModel (HMM)
a space filling curve allows to approximate the 2D as a simple chain
Lazrak2012, PhD thesis
biblioat http://bit.ly/IALE2014
9. Methods#4 | hierarchicalmodel
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 10
HierarchicalHMM isstrictlyergodicto segment the landscapeintopatches of homogeneoustime sequences
Each “super state” of a “super model” is a HMM (generally linear) to describe the (time) sequences
Twostochasticprocessescan be represented,
the first (hidden) driving
the second(knwonand visible)
HMM= Hidden Markov Model
Combiningheterogeneousdata to deal with drivers of management dynamics(insteadof pixels)
ArpentAgeinstead of clustering because “context- wise” (cf. fractal curve)
Context-wiselandscapesegmentationfocusedon management relatedvariables
10. Materials#1 | studyarea: Monte Pisano
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 12
Mediterranean terraced olive grove farming system (62 km2)
11. Materials#2 | input data
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 13
Markov random field approximated to a hidden Markov model using a space-filling curve
Layers were sampled on a 10m regular point grid
Geology
5 cl.
• anthropic • alluvial & sand-loam sediments • quartzes • limestone • schist
Orientation
4 cl.
N, S, E, W
Morphology
4 cl.
• Plains • valleys • slopes • ridges
Land cover
14 cl.
• Urban • garden
• cultivated olive groves • abandoned olive groves • vineyards • orchards • arable lands • fallows • garigues• outcrops • wetland vegetation • water • pine wood • wood
Terracetype
5 cl.
(bench) terraces • talwegs• pocket terraces • broad field terraces • extensive
Dry-stonewalldensity
6 cl.
From 0to 44.5 m/ha (natural breaks)
12. Results#1 | an output map
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 15
terraces
talwegs
extensive
West exposedterraces
periurban
warmwoods
coldwoods
terracedgroves
foothills
hectares
13. Results#2 | variablesegmentation
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 16
Three majorsmanagement regions:
• Terracedgrovesmedium-high densityDSW on quartz, E-S orientedwith some gariguesand pine woods
• Woodsmaydifferfor morphologyor exposition
• Foothillsmaybe segmentedfor specificterracetypes, mix with urbanlandcover or plainmorphology
The increasing segmentation allows to define a hierarchy of the variables
Especially relevant for management issues
periurban
periurban
14. Results#3 | location of abandonedolive groves
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 17
Eastern side
prevalently associated to sand-loam sediment area, with low density terraced in valleys.
A
C
B
A
B
C
Western side
prevalently associated to the cluster grouping the foothills, with the prevalence of the anthropic bedrock and S-W exposition
West exposedterraces
periurban
terracedgroves
foothills
15. Conclusionsand perspectives
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 18
Flexibilityof variables treated remote sensing map, census data, etc.
Flexibilityof scale of analysis potentially no limit if input data are defined coherently to the research question
variable choicee.g., what if we introduce proximity to road?
input variable classificationtype and number of classes
variable input qualitydifferent geometric scales
HHMM parameterizationspacefillingcurve, numberof states
Future applications could address the temporospatialcharacterization of agricultural practicesat the watershed scale
Leverages…
…and potentials
16. Participate!
9th IALE World Congress
July 5-10, 2015
Portland, Oregon
Landscape agronomy: advances and challenges of a new field addressing the management of rural land systems
read more at
http://bit.ly/IALE2015_Symposium_LA
RIZZO Davide,
FERGUSON Rafter Sass,
BENOIT Marc,
PADOA-SCHIOPPA Emilio
contact & info ridavide@gmail.com
IALE-Europe Thematic Workshop 4th July 2014 Agricultural landscape segmentation • Rizzo et al. • Symposium 1 19