[Defended thesis] Arthur Crespin-Boucaud: Remote sensing and knowledge integration through spatial modelling for a more consistent mapping of complex agricultural systems: Application to the Highlands region in Madagascar

Arthur is one of the #DigitAg labelled PhDs

 

Arthur defended his thesis on 23rd March at 9 AM.

Remote sensing and knowledge integration through spatial modelling for a more consistent mapping of complex agricultural systems: Application to the Highlands region in Madagascar

  • Start Date: October 2017
  • University: AgroParisTech
  • PhD School: GAIA
  • Field(s): Geomatics, Remote sensing, Spatial Modeling
  • Doctoral Thesis Advisor: Agnès Bégué, UMR Tetis, Cirad
  • Co-supervisors : Valentine Lebourgeois, Danny Lo-Seen, Mathieu Castets  (UMR Tetis, Cirad)
  • Funding: Cirad – CNES
  • Linked Project: SAMSAM (Satellite time-series Analysus and Modeling for Small Agriculture Mapping)
  • #DigitAg: Axis 6 – Challenge 6

 

Keywords: Remote sensing, Spatio-temporal modeling, Family farming, Land cover and land use mapping

Abstract: Classification methods for remote sensing data for land cover mapping are based on proximity measurements between pixels or objects in a spectral hyperspace in which they are projected according to their reflectance in several spectral bands of the image.

Then projecting the pixels or objects from the hyperspace back into geographical space yields the land cover map sought, with as many land cover classes as regions identified as meaningful in the hyperspace. The rationale behind the analysis within the satellite data hyperspace is that in the latter, relationships involving the “spectral signatures” of objects and pixels are easier to retrieve. Such methods have proved very efficient in mapping agricultural landscapes in many parts of the world characterised by large, well-managed intensive agricultural plots.

However, for large parts of the developing world, agricultural landscapes are more complex, with typically smaller plots, and a diversity of land use and land cover that reflect local customs and their adaptation to local climate and geography. The efficiency of remote sensing methods sharply decrease for the more complex landscapes. In order to push back these limits, current researches are focusing on methods combining both time series of satellite images and contextual data, such as texture indices, altitude or slope, some of which are based on deep learning.

These methodological developments and in-depth techniques use mainly spectral information and integrate little of the other types of agricultural knowledge available. For example, it is known that some crops only grow above a given altitude or close to households. These types of (spatial and temporal) relationships are easy to formalise in geographical space, but much less so in the data hyperspace. This intuitively suggests that better taking into account agricultural knowledge could improve classification methods to obtain agriculturally consistent land use maps.

In this thesis, we explore the possibility of using agricultural knowledge, formalised as rules to improve an existing machine learning classification method for land cover and land use mapping in complex agricultural systems. First, this thesis proposes a conceptual model allowing to combine both remote sensing  methods and spatio-temporal modelling. This model is decomposed into four spatial and temporal modules, each corresponding to a method aimed at improving the characterization of land cover and land use, and which can be used independently. The two spatial modules of the model are then applied to an agricultural study area located in the Vakinankaratra region, in the highlands of Madagascar, in order to evaluate the approach developed.

In quantitative terms, the application of the two spatial modules only slightly improves the characterisation of the land use classes of the study area, mainly due to the lack of quality data supporting the application of the spatial rules. Nevertheless, the application of spatial modules allows a better discrimination between rainfed crops and savannah areas, which is a source of much confusion with the methods used in remote sensing. The analysis of these results leads to improvements for the conceptual model as well as for its more general application to complex agricultural systems.

Contact:  arthur.crespin-boucaud [AT] cirad.fr​ – +33 467558645


International papers

Crespin-Boucaud, A.; Lebourgeois, V.; Lo Seen, D.; Castets, M.; Bégué, A. Agriculturally consistent mappingof smallholder farming systems using remote sensing and spatial modelling, ISPRS—Int. Arch. Photogramm.Remote Sens. Spat. Inf. Sci.2020,XLII-3/W11, 35–42

Bélières   Jean-François   (dir.),   Dianah   Randriamitantsoa,   Harimandranto   Randrianirina,Noroseheno Ralisoa, Arthur Crespin-Boucaud, 2020, Étude chaîne de valeur pomme de terre.Partie 1: Importance de la culture de la pomme de terre pour les exploitations agricoles etrentabilité de la production de plants de semence et de consommations.  CASEF,  MAEP