Offre en lien avec l’Action/le Réseau : MACLEAN/– — –
Laboratoire/Entreprise : TETIS
Durée : 3 ans
Contact : raffaele.gaetano@cirad.fr
Date limite de publication : 2020-06-15
Contexte :
Extensive farming systems, still widespread in the tropics, are generally based on fallow practices, because of their ability to regenerate soil fertility, particularly through the maintenance of biomass reservoirs. Their importance has also been emphasized in adaptation to climate change, as they contribute to carbon sequestration and the reduction of greenhouse gases [1]. As a result, the estimation of fallow areas is an important piece of information in assessing the performance of an agricultural system, both in terms of short-term productivity and the quantification of the “land stock” available for the establishment of strategies in response to climatic and/or anthropogenic factors. If the documentation of this practice in different regions of the world is important, a regular and exhaustive inventory of fallow land in West Africa, which would allow to better study their spatio-temporal dynamics (extent and duration, farmers’ strategies, the role of climate) does not exist [2, 3]. Given the stakes involved in this practice, this thesis aims to define a methodological framework combining expert knowledge and satellite imagery for the implementation of a fallow monitoring system at large scale. Indeed, fallow mapping is poorly taken into account in real-world remote sensing based land cover products. At best, this problem is naively approached [4] omitting any consideration of the specificities related to these practices (extent, duration, strategies, but also their role in landscapes). In order to overcome these limitations, and with a growing availability of satellite images adapted to the monitoring of West African complex agricultural landscapes (such as those from ESA’s Sentinel missions), we will promote an interdisciplinary approach to (i) study the relationship between fallow land use and remote sensing indicators and (ii) to convey this information in the design of methods for analyzing multi-year series of images for their identification and characterization. More precisely, we rely on the hypothesis that one of the main issues that limit the possibilities in terms of automatic mapping of fallow practices, beyond those related to lack of sufficient reference data, is the fact that fallows are mainly modeled as a specific land use class. Matter of facts, other factors such as the age and/or human intervention can have a significant impact on its radiometric response. To cope with these limitations, we are going to promote alternative approaches based on modeling land use, which concerns the anthropic management of surfaces, in terms of “land cover trajectories” at larger time scales, which on the contrary focus on the dynamics of the biophysical covers observed over surfaces [5]. To this aim, we are going to mobilize, in complement to the seasonal spectral information, the multi-year information coming from high-resolution optical and radar (Sentinel-1 and -2) imagery, to address the following specific objectives:
• develop a methodology for the automatic land cover mapping based on supervised classification of multi-sensor image time series, being capable to provide a spatial information which is consistent and comparable through multiple years (i.e., robust with respect to non-stationary seasonal specificities – climatic, related to landscape changes, but also to image availability). This part will focus on specific research on machine learning topics such as domain adaptation and optimal transfer [6], as well as on the availability of a reference data base available over an agricultural area in the South-West of Burkina Faso (Koumbia, province of Tuy), covering a period between 2013-2018;
• identify novel strategies allowing the deployment of our approach in cases where only reference data on a single year are available, based on the extraction of variables directly issued by multi- year image time series, whose efficiency will be tested on other agricultural systems with close (groundnut basin in Senegal) or contrasted (slash-and-burn practices in Guinea) fallow practices.
This part of the work will be based on coupling expert modeling of land cover changes with data- driven approaches, notably those based on the use of novel machine and deep learning techniques adapted to image time series [7].
Sujet :
Extensive farming systems, still widespread in the tropics, are generally based on fallow practices, because of their ability to regenerate soil fertility, particularly through the maintenance of biomass reservoirs. Their importance has also been emphasized in adaptation to climate change, as they contribute to carbon sequestration and the reduction of greenhouse gases [1]. As a result, the estimation of fallow areas is an important piece of information in assessing the performance of an agricultural system, both in terms of short-term productivity and the quantification of the “land stock” available for the establishment of strategies in response to climatic and/or anthropogenic factors. If the documentation of this practice in different regions of the world is important, a regular and exhaustive inventory of fallow land in West Africa, which would allow to better study their spatio-temporal dynamics (extent and duration, farmers’ strategies, the role of climate) does not exist [2, 3]. Given the stakes involved in this practice, this thesis aims to define a methodological framework combining expert knowledge and satellite imagery for the implementation of a fallow monitoring system at large scale. Indeed, fallow mapping is poorly taken into account in real-world remote sensing based land cover products. At best, this problem is naively approached [4] omitting any consideration of the specificities related to these practices (extent, duration, strategies, but also their role in landscapes). In order to overcome these limitations, and with a growing availability of satellite images adapted to the monitoring of West African complex agricultural landscapes (such as those from ESA’s Sentinel missions), we will promote an interdisciplinary approach to (i) study the relationship between fallow land use and remote sensing indicators and (ii) to convey this information in the design of methods for analyzing multi-year series of images for their identification and characterization. More precisely, we rely on the hypothesis that one of the main issues that limit the possibilities in terms of automatic mapping of fallow practices, beyond those related to lack of sufficient reference data, is the fact that fallows are mainly modeled as a specific land use class. Matter of facts, other factors such as the age and/or human intervention can have a significant impact on its radiometric response. To cope with these limitations, we are going to promote alternative approaches based on modeling land use, which concerns the anthropic management of surfaces, in terms of “land cover trajectories” at larger time scales, which on the contrary focus on the dynamics of the biophysical covers observed over surfaces [5]. To this aim, we are going to mobilize, in complement to the seasonal spectral information, the multi-year information coming from high-resolution optical and radar (Sentinel-1 and -2) imagery, to address the following specific objectives:
• develop a methodology for the automatic land cover mapping based on supervised classification of multi-sensor image time series, being capable to provide a spatial information which is consistent and comparable through multiple years (i.e., robust with respect to non-stationary seasonal specificities – climatic, related to landscape changes, but also to image availability). This part will focus on specific research on machine learning topics such as domain adaptation and optimal transfer [6], as well as on the availability of a reference data base available over an agricultural area in the South-West of Burkina Faso (Koumbia, province of Tuy), covering a period between 2013-2018;
• identify novel strategies allowing the deployment of our approach in cases where only reference data on a single year are available, based on the extraction of variables directly issued by multi- year image time series, whose efficiency will be tested on other agricultural systems with close (groundnut basin in Senegal) or contrasted (slash-and-burn practices in Guinea) fallow practices.
This part of the work will be based on coupling expert modeling of land cover changes with data- driven approaches, notably those based on the use of novel machine and deep learning techniques adapted to image time series [7].
Profil du candidat :
Prerequisites : the PhD thesis subject mainly require skills on methodologies for remote sensing image processing and analysis, as well as the related software development tools (including machine learning). An experience on the application of remote sensing imagery to agriculture will be strongly appreciated.
• Training profile : students coming from any course providing solid bases in remote sensing as well as its applications in agricultural, environmental, and/or territorial monitoring and management. Candidates coming from study courses much related to data science and artificial intelligence, having an experience with the use of remote sensing data, will also be taken into consideration.
• Work environment : the PhD student will mainly work with software development tools (Python language programming – preferably, machine learning packages and libraries for remote sensing image processing – e.g. the Orfeo Toolbox). Field missions on West-African study sites are possible during the PhD thesis duration.
• Candidates must have a good level on written English (for the production of scientific papers in international conferences and journals), and a sufficient spoken skill to be autonomous in the exchanges with foreign partners.
Formation et compétences requises :
Prerequisites : the PhD thesis subject mainly require skills on methodologies for remote sensing image processing and analysis, as well as the related software development tools (including machine learning). An experience on the application of remote sensing imagery to agriculture will be strongly appreciated.
• Training profile : students coming from any course providing solid bases in remote sensing as well as its applications in agricultural, environmental, and/or territorial monitoring and management. Candidates coming from study courses much related to data science and artificial intelligence, having an experience with the use of remote sensing data, will also be taken into consideration.
• Work environment : the PhD student will mainly work with software development tools (Python language programming – preferably, machine learning packages and libraries for remote sensing image processing – e.g. the Orfeo Toolbox). Field missions on West-African study sites are possible during the PhD thesis duration.
• Candidates must have a good level on written English (for the production of scientific papers in international conferences and journals), and a sufficient spoken skill to be autonomous in the exchanges with foreign partners.
Adresse d’emploi :
UMR TETIS,
500 Rue Jean-François Breton, Montpellier
Document attaché : 202005151313_DIGITAG_TheseJachere_CIRAD-TETIS.pdf